HyperFrames Deploys Claude to Build Complete Videos from Screenshots

How to Make Amazing Videos with Claude and HyperFrames | Jake Moran — Peter Yang

Peter Yang

AI video tools can automate timeline assembly, jump cuts, and style replication, shifting human production work entirely to conceptual direction, brand constraints, and frame-level annotations.

AI-assisted summary of How to Make Amazing Videos with Claude and HyperFrames | Jake Moran — original source. Check the original for full context.

Source published:

Key ideas

  • HyperFrames leverages an HTML-based architecture that enables code-native AI agents like Claude to programmatically assemble product launch videos and rough cuts from raw footage and static screenshots. (0:00)
  • Sub-second video trimming requires precise audio alignment; running NVIDIA's local 1 GB Parakeet speech-to-text model delivers the millisecond timestamp accuracy needed to prune pauses down to 250 milliseconds without clipping syllables. (6:14)
  • Product launch videos achieve optimal viewer conversion by following a four-beat structure: the hook, context building, the payoff or call to action, and the brand reminder. (20:00)
  • HyperFrames simplifies editing iteration by allowing users to draw feedback directly onto individual video frames, queuing visual corrections and notes for batch submission to Claude. (36:17)

Detailed notes

## Background Published by Peter Yang on YouTube on October 25, 2026, this episode features AI video expert Jake Moran explaining how to prompt Claude and HyperFrames to produce high-impact videos. The discussion focuses on workflows for product launches, long-form YouTube tutorials, and short-form social clips, detailing the core editorial terminology needed to direct generative models effectively.

## TL;DR Jake Moran outlines an end-to-end framework for AI-assisted video editing, combining reasoning models like Claude with HyperFrames and NVIDIA's Parakeet for millisecond-precise rough cuts. He demonstrates how to transform static screenshots into polished motion graphic launch videos using a four-part narrative structure, integrate synthetic HeyGen avatars, and automatically adapt brand palettes. The conversation covers advanced techniques for cloning visual pacing from short reference clips, utilizing in-frame visual drawing canvases for batch feedback, and segmenting long-form content to preserve token limits.

## Atmosphere & style The conversation is pragmatic, collegial, and deeply builder-focused, pairing Peter Yang's eager questions as a non-editor creator with Jake Moran's relaxed, technical fluency. It feels like an intimate screen-share workshop where practical editorial concepts—like whip pans, punch-ins, and easing curves—demystify what typically seems out of reach for non-specialists. The mood is energized by genuine surprise and excitement, especially when observing Claude make unprompted creative decisions like inserting Jake's digital twin or converting five-second reference videos into reusable programmatic templates.

## Bulletin - [00:00](https://www.youtube.com/watch?v=MC1g2fHYC88&t=0s) HyperFrames leverages an HTML-based architecture that enables code-native AI agents like Claude to programmatically assemble product launch videos and rough cuts from raw footage and static screenshots. - [06:14](https://www.youtube.com/watch?v=MC1g2fHYC88&t=374s) Sub-second video trimming requires precise audio alignment; running NVIDIA's local 1 GB Parakeet speech-to-text model delivers the millisecond timestamp accuracy needed to prune pauses down to 250 milliseconds without clipping syllables. - [20:00](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1200s) Product launch videos achieve optimal viewer conversion by following a four-beat structure: the hook, context building, the payoff or call to action, and the brand reminder. - [36:17](https://www.youtube.com/watch?v=MC1g2fHYC88&t=2177s) HyperFrames simplifies editing iteration by allowing users to draw feedback directly onto individual video frames, queuing visual corrections and notes for batch submission to Claude.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=MC1g2fHYC88&t=0s) Effective prompt-driven video editing relies on using standard editorial terminology—such as rough cuts, jump cuts, and punch-ins—to give the LLM clear structural context. - [02:30](https://www.youtube.com/watch?v=MC1g2fHYC88&t=150s) To achieve a tight rough cut, users should provide explicit thresholds, such as cutting thinking pauses to 0.25 seconds or less and dropping false starts to keep only final takes. - [04:45](https://www.youtube.com/watch?v=MC1g2fHYC88&t=285s) Punch-ins and alternating camera framings are used to mask the head-position shifts caused by jump cuts when silences or retakes are removed from a single continuous take. - [06:14](https://www.youtube.com/watch?v=MC1g2fHYC88&t=374s) Transcription drift is a major failure mode in automated video editing; low-latency local models like NVIDIA's Parakeet solve this by providing accurate word-level timestamps. - [08:40](https://www.youtube.com/watch?v=MC1g2fHYC88&t=520s) Camera zooms—whether continuous push-ins or abrupt cut-ins—serve as essential visual hooks within the opening seconds to draw viewers into the scene. - [10:00](https://www.youtube.com/watch?v=MC1g2fHYC88&t=600s) Because LLMs lack auditory perception to evaluate sound balance, prompts must explicitly set background audio ratios, such as keeping music volume at 10% of the spoken dialogue. - [13:37](https://www.youtube.com/watch?v=MC1g2fHYC88&t=817s) HyperFrames uses a storyboard mode to generate static keyframes first, allowing creators to inspect layout, composition, and tone before committing compute to a full motion render. - [20:39](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1239s) Modern generative video workflows can fully synthesize on-screen talent by integrating digital avatars from tools like HeyGen directly via desktop app bridges or MCP connectors. - [22:20](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1340s) Brand consistency requires explicit input; HyperFrames extracts color palettes and styling directly from website URLs or style guides to prevent AI agents from hallucinating random aesthetics. - [26:40](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1600s) Pacing requirements differ fundamentally across platforms: short-form feeds demand constant kinetic cuts, whereas long-form educational videos succeed with an upfront table of contents and breathing room for the speaker. - [30:00](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1800s) Five-second reference clips can be reverse-engineered by AI into underlying layout and motion code, enabling creators to clone dynamic editing styles while replacing the branding and assets. - [36:17](https://www.youtube.com/watch?v=MC1g2fHYC88&t=2177s) To prevent excessive token usage and slowdowns on videos longer than 30 minutes, editors should chunk the timeline into modular chapters rather than processing entire long-form recordings in a single pass. - [36:17](https://www.youtube.com/watch?v=MC1g2fHYC88&t=2177s) Successful compound prompts and workflow pipelines should be saved as reusable agent skills to maintain consistent editing formats across future video productions.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=MC1g2fHYC88&t=0s) Teaser, Intro, and Video Editing 101 with AI

The conversation opens with a glimpse of an impressive AI-generated video created from minimal input: three simple screenshots that Claude expanded into a full video, populating it with synthetic people and even inserting Jake Moran's own AI avatar without being prompted. As Jake points out, Claude took the creative liberty entirely on its own: “That's my avatar as well. Yeah, and I didn't even ask for myself to be in there. That was actually Claude's discretion.”

Peter Yang introduces Jake as his go-to resource for AI video editing. Despite having edited videos for only six months, Jake has developed an efficient methodology using AI agents to generate product launch videos, talking-head YouTube episodes, and short-form content. Peter confesses that he is a complete novice when it comes to video editing terminology and asks Jake to establish the foundational concepts and vocabulary needed to instruct AI agents effectively.

Jake explains that video creation spans many formats—from marketing launch videos to long-form recorded podcasts that require extensive post-production cleanup. When a recording session finishes, the raw, uncut footage (the "raw takes" or rushes) is invariably longer, looser, and messier than the final intended product. AI models excel at compressing that raw footage into a tight, polished delivery, provided the user communicates in terms that the models understand natively.

```mermaid flowchart TD accTitle: Raw Footage Cleanup Flow accDescr: Pipeline showing how raw recorded takes are processed by an LLM agent into a polished cut using transcription and editing directives.

subgraph Ingestion ["1. Raw Ingestion"] A["Raw Takes / Rushes"] --> B["Transcription Model\n(Timestamps & Words)"] end

subgraph AgentPass ["2. LLM / Agent Editing Pass"] B --> C["Identify Artifacts\n(Filler Words, Pauses, Retakes)"] C --> D["Apply Directives\n(Cut pauses to <0.25s, drop duplicate takes)"] end

subgraph Assembly ["3. Visual Assembly"] D --> E["Rough Cut\n(Storyline Restored)"] E --> F["Pacing & Transitions\n(Jump Cuts, Punch-ins, Framing)"] end ```

- HyperFrames relies on an HTML-based architecture, which makes it particularly natural for code- and web-native AI agents to manipulate programmatically compared to legacy timeline software. - Speaking the standard editorial vocabulary ensures the LLM recognizes the exact intent immediately, preventing wasted tokens and expensive trial-and-error iterations.

---

[02:30](https://www.youtube.com/watch?v=MC1g2fHYC88&t=150s) Mastering the Rough Cut: Filler Words, Pauses, and Retakes

Jake breaks down the mechanics of assembling a "rough cut"—the critical early stage where mistakes, stumbles, and dead air are eliminated to re-establish a clear, engaging narrative arc. Because modern LLMs operate over detailed transcripts paired with accurate word-level and sentence-level timestamps, they can effortlessly identify common speech quirks such as verbal pauses, stutters, and duplicate attempts at saying the same line.

To get clean results on the first attempt, Jake emphasizes that vague prompts yield messy outputs. Instead, creators should provide granular instructions and exact thresholds. Users should explicitly tell the agent to strip out verbal crutches like "um" and "uh," trim long thinking pauses down to a tight duration (such as a quarter of a second or less), discard false starts, and retain only the final take when a speaker blunders and repeats themselves.

Jake stresses that context is the single most decisive factor when delegating video work to an LLM: “Basically, I think this is true with all things AI, but it's also especially true with video editing. Getting as much context into your prompt, into wherever you're asking things to be edited is the key to increasing quality and getting closer to what you're wanting.”

- **Filler Words:** Removing hesitations ("um," "ah," "like") that distract from the core argument. - **Micro-Pauses:** Pruning silence between phrases to roughly 0.25 seconds or less to maintain high-energy pacing. - **Retakes and Blunders:** Instructing the model to recognize repeated phrasing, drop the aborted attempt, and splice directly into the corrected delivery.

---

[04:45](https://www.youtube.com/watch?v=MC1g2fHYC88&t=285s) Jump Cuts, Punch-Ins, and Modern LLM Capabilities

When silence or duplicate words are cut out of a continuous camera take, the speaker's head will instantly shift position between frames. Jake explains that this visual jump is known as a jump cut. To make these jarring edits feel deliberate and visually dynamic rather than accidental, editors use "punch-ins"—zooming into the shot slightly or cutting between closer framing and wider framing to mask the stitch. Similarly, punching in can highlight an emphasized statement or introduce a visual shift that re-engages the viewer's focus.

Peter reflects on how tedious this used to be when done by hand, recalling how creators had to sit and manually cross-reference the video timeline against a text transcript to slice out dead air. He observes that recent frontier models have completely altered the speed of this workflow, allowing users to combine all these cleanup instructions—cutting filler words, eliminating repeated lines, applying jump cuts, and punching in—into a single compound prompt.

Jake agrees that this rapid automation has only become dependable over the preceding few months, driven by advances in high-end reasoning and multimodal models like Claude 3.5 Sonnet, OpenAI models, and Google's Gemini lineup, which can ingest large transcripts, handle spatial/temporal reasoning, and execute multi-part editing instructions in a single shot.

---

[06:14](https://www.youtube.com/watch?v=MC1g2fHYC88&t=374s) Parakeet: Accurate Local Transcription for Millisecond Cuts

While top-tier LLMs excel at creative and structural editing logic, their cuts depend entirely on the quality of the underlying audio timestamps. Jake highlights a hidden bottleneck: if a transcription engine provides sloppy or drift-prone timestamps, the agent will inevitably misjudge the audio boundaries and chop off the beginning or end of words.

To solve this, Jake relies on NVIDIA's Parakeet model, a locally hosted speech-to-text model that is approximately 1 GB in size. For short-form platforms like TikTok, YouTube Shorts, or Instagram Reels where breathless pacing is mandatory and pauses are shaved down to 250 milliseconds, microsecond precision is the difference between an edit that feels seamless and one that sounds broken.

Jake explains why transcription accuracy is the bedrock of automated editing: “So Parakeet for me works 98% of the time to getting the right amount of time in between words or whatever the case you need. It just allows the LLM to get really good timestamps, which is everything for these types of, let's say, recording cleanup.”

- **The Precision Problem:** Low-quality automated transcription generates loose timestamps; asking an LLM to cut tight silences against inaccurate timestamps clips the first or last syllable of neighboring words. - **Local Model Footprint:** At roughly 1 GB, Parakeet runs locally, offering low latency, zero external API costs, and high-fidelity timing for video editing pipelines. - **Pacing Demands:** Social video algorithms punish sluggish dead air, making sub-second cuts vital for viewer retention.

---

[08:00](https://www.youtube.com/watch?v=MC1g2fHYC88&t=480s) Sponsor Break: Wispr Flow

Peter shares a brief sponsorship message for Wispr Flow, an AI-powered voice dictation application that saves him at least three hours every week. Instead of typing out thoughts, users speak naturally, and the app automatically removes filler words, formats the sentences, and outputs clean, ready-to-use prose. Peter uses Wispr Flow across his daily workflow to draft newsletter editions, write detailed product specs, and answer messages on Slack, offering listeners six months free with his promo code.

---

[08:40](https://www.youtube.com/watch?v=MC1g2fHYC88&t=520s) Crafting the Hook: Zooms, Push-Ins, and Cut-Ins

Turning from cleanup to creative presentation, Peter asks Jake how to tackle the intro of a video. Jake notes that the terms "intro" and "hook" are often used interchangeably across the video production landscape. Regardless of what label an editor prefers, the fundamental objective remains identical: capture the audience's visual and intellectual attention within the first few seconds before they swipe away.

When working with talking-head footage, simple camera movement serves as one of the most reliable visual hooks available. Jake stresses the psychological impact of moving the frame: “I can't understate how powerful a zoom in is at the beginning of the video. It just leads your eye somewhere and kind of wraps you into the world.” Modern LLMs can be prompted to script and automate these zooms without manual keyframing.

Jake distinguishes between the primary zoom techniques editors use to frame an opening: - **Push-Ins:** A gradual, continuous camera motion that starts wider and steadily pushes forward closer to the subject's face, gently pulling the viewer into the scene. - **Cut-Ins:** An abrupt change in focal length where the camera cuts directly from a wide angle to a tight shot on a specific frame without any visible continuous zoom motion.

[10:00](https://www.youtube.com/watch?v=MC1g2fHYC88&t=600s) Hooks, Dynamic Transitions, and AI-Driven Sound Design

Jake Moran explains that dynamic camera and frame movements in video editing rely heavily on pacing curves. Rather than jarringly jumping between position frames, most creators utilize an easing curve that starts slow, ramps up in speed through the middle, and smooths out toward the end. Beyond camera motion, intros require B-roll to capture attention immediately. B-roll can serve multiple purposes: teasing topics discussed later in the video, presenting visually engaging eye candy, displaying a tutorial's table of contents, or showing speaker pedigree cards. Moran points out that established creators like Peter Yang often use pedigree overlays to signal authority and history to first-time viewers right away.

Transition choices further define a video's energy. Moran contrasts a standard cut (an abrupt swap from scene A to scene B) with a whip pan, where scene A snaps rapidly off-screen to the left while scene B whips in from the right. When Yang admits he does not always know when to choose a whip pan over a cut and asks if one can simply prompt Claude to pick from a list, Moran strongly encourages letting the LLM take the reins:

> “Leave it up to Claude, it's going to make the choice for you, and it'll probably be pretty solid.”

When it comes to audio, Claude 3 Opus can generate its own MIDI music tracks, hook into audio libraries, or accept local music files dropped directly into the chat prompt. However, Moran cautions that language models cannot actually listen to audio, making balance a common failure mode where background music overpowers the dialogue. To prevent this, he explicitly instructs the agent to keep music levels at 10% of the A-roll voiceover volume. Yang adds that prompting the AI to snap cuts directly to the beat of the music elevates the edit to a professional level.

- **Easing Curves:** Motion graphics feel natural when movement accelerates in the middle and softens at the endpoints rather than moving linearly. - **B-Roll Varieties:** Used for visual pacing, previewing upcoming segments, providing structured tables of contents, or establishing social proof with credential overlays. - **Whip Pans:** Directional blur transitions that maintain high kinetic energy between disparate scenes. - **Audio Prompting:** Always specify explicit volume ratios (e.g., “keep it at 10% of the volume of my A-roll”) because LLMs have no auditory perception to judge loudness balance automatically.

---

[13:37](https://www.youtube.com/watch?v=MC1g2fHYC88&t=817s) Demo: Building a Feature Launch Video from Three Screenshots

Moran walks through a concrete demo showing how he produced a launch video for HyperFrames' native background-removal tool. Rather than assembling complex timelines manually, he built the entire motion graphic sequence inside the HyperFrames desktop app starting from just three static assets: a screenshot of the app's studio interface and two reference images for the hook. Instead of specifying millisecond durations, cut points, or technical editing jargon, he wrote a high-level narrative prompt focused on the product storyline, relying on Claude to handle the graphic composition and pacing.

```mermaid flowchart TD accTitle: AI Video Creation from Assets to Render accDescr: Workflow showing how three screenshots and a story idea turn into a storyboard and final video build via Claude.

subgraph Inputs ["1. Input Assets & Story"] A["3 Static Screenshots"] B["Core Storyline & Slack Hook Asset"] end

subgraph Prompting ["2. LLM Orchestration"] C["High-Level Narrative Prompt<br/>(No complex jargon)"] D["Claude 3 Opus Agent"] end

subgraph Iteration ["3. Review & Render"] E["Still-Frame Storyboard"] F["Targeted Prompt Adjustments"] G["Full Motion Graphics Render"] end

A --> C B --> C C --> D D --> E E -->|"Inspect & refine frames"| F F --> D E -->|"Approved"| G ```

Moran highlights a fundamental truth about AI-assisted video editing: the creator still carries the burden of taste, conceptual direction, and the core premise:

> “The biggest bummer I think for people still today with editing with AI is that you actually, you probably do still need your idea, right?”

In this instance, inspiration struck when an engineer on Moran's team posted a Slack demo featuring an image of a basketball player with the background cleanly removed. Moran grabbed that exact asset, dropped it into the chat, and drafted a script conveying that complex background extraction was now "just a prompt away."

To avoid the lengthy wait times of rendering full motion passes, Moran leverages HyperFrames' storyboard mode. The model generates static keyframes representing the video's chronological beats first. This review layer allows him to evaluate layout, composition, and visual tone, requesting style adjustments on still frames before burning computational time on a full video render.

---

[18:41](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1121s) Guidelines for Fast-Turnaround Feature Launch Videos

Yang asks Moran for his general framework when producing product launch videos, which typically run between 30 and 45 seconds to highlight a problem and its resolution. Moran clarifies that his launch videos are rarely sweeping, documentary-style founder stories or grandiose funding announcements. Instead, they are rapid-response feature releases tailored for specific user workflows.

At an early-stage startup, engineering moves faster than conventional video production schedules. Moran routinely receives Slack messages at 5:00 PM from engineers saying a feature is completed and asking if they can announce it the next morning. Because the target audience only needs to understand a focused utility, Moran aims for concise videos running between 30 seconds and roughly a minute. Every video opens with an immediate, visually gripping hook—such as the basketball visual used in the background removal demo—before efficiently demonstrating the specific workflow improvement.

[20:00](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1200s) Structuring Launch Videos into Four Core Beats

Jake breaks down the mental framework he uses when structuring product launch videos. He thinks about them in four essential sections: the hook, context building, the payoff or call to action (CTA), and the brand reminder. The hook needs to be the single most engaging element that grabs the viewer's attention right away. From there, the video transitions into explaining how the product actually works in context, followed by demonstrating the ultimate payoff for the end user. Finally, the video loops back around to pricing or practical details, reminding the audience of who made the tool:

> "I normally think about it in four sections, hook, kind of like the context building, maybe the call to action or the output, and then coming back around to like, hey, here's our brand."

```mermaid flowchart LR accTitle: Four-Part Product Launch Video Framework accDescr: A linear four-stage workflow showing how Jake structures product launch videos.

A["1. The Hook<br/>(Most engaging element)"] --> B["2. Context Building<br/>(How to use the product)"] B --> C["3. The Payoff / CTA<br/>(Value to user & outcome)"] C --> D["4. Brand Reminder<br/>(Pricing, details & brand identity)"] ```

[20:39](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1239s) Using HeyGen Avatars and Claude's Unprompted Decision

Peter points out a video clip showing a massive grid wall of talking avatars and asks whether Claude generated the human figures directly or if Jake used HeyGen. Jake explains that as a HeyGen employee who was with the company long before starting HyperFrames, he prefers not being on camera and lacks direct access to an army of creators. Rather than spending hours scouring TikTok or Instagram for sample footage, he simply asked Claude to generate different digital personas through HeyGen:

> "In this case like we see later in this video right we have this crazy grid wall of video examples I could have pulled these from online I lazy And I just asked Claude to use HeyGen to generate different characters and then have them speak a couple clips."

When Peter asks if anyone in the clip is real—even noticing Jake's likeness in the center—Jake reveals that every single person, including himself, is an avatar. Intriguingly, Jake had not even prompted Claude to insert his personal digital twin:

> "Yeah, and I didn't even ask for myself to be in there. That was actually Claude's discretion."

Jake notes that if a creator already has raw footage, they can just drop it in. To connect models without leaving the editing environment, HyperFrames incorporates direct integrations: - The HyperFrames desktop app features native direct connections to generative AI models. - When running workflows through the Claude desktop application or code editors like Codex, users connect external generators (HeyGen for avatars, Suno for audio, or various image generation tools) using CLI or Model Context Protocol (MCP) servers.

```mermaid flowchart TD accTitle: External Generative Tools Integration accDescr: Diagram showing how external tools like HeyGen and Suno integrate into editing workflows via app or MCP.

subgraph Native["HyperFrames Desktop App"] Direct["Direct Native Integrations"] end

subgraph External["Third-Party AI Editors (Claude Desktop / Codex)"] Bridge["CLI / MCP Connector"] end

Direct --> Models["Generative Models<br/>(HeyGen avatars, Suno audio, ImageGen)"] Bridge --> Models ```

[22:20](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1340s) Maintaining Visual Brand Consistency

Peter asks how creators can guarantee that AI agents keep brand colors, fonts, and styling consistent from one cut to another instead of hallucinating random aesthetics. Jake highlights a feature they built into HyperFrames that automatically derives an entire brand kit directly from a user's website. Alternatively, users can upload a brand guidelines PDF or an asset folder.

Jake cautions that failing to supply concrete brand context leads to unpredictable results:

> "If you just say my brand, probably nine times out of 10, you're going to get some crazy colors and an imagined surface."

Inside HyperFrames, brand kit detection happens natively, whereas creators working in generalist LLM setups must explicitly feed style guides and color codes into the prompt context.

[23:28](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1408s) Editing Peter's Talking-Head Video Down from Raw Footage

The conversation shifts from short launch promos to long-form tutorials. Jake reveals that he took a raw, unedited 20-minute and 30-second episode recorded by Peter and put it through the AI pipeline. The resulting polished video came out to 18 minutes and 11 seconds.

Jake explains that the 2-minute-and-19-second reduction stemmed from using an automated transcription pass to strip out pauses, duplicate takes, and verbal flubs. Depending on how messy the unscripted recording is, the cut-down ratio can vary widely:

> "Maybe I've recorded for eight minutes before and because I was sloppy and didn't have a script and whatever, the finished output of what I actually needed was only a minute, right? That's that you never really know it just depends on how let's say clean you were through the creation process there"

Beyond pruning filler, Jake focused on building an engaging intro hook reel that introduced the video's core topics.

[25:00](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1500s) Building the Intro Reel with Screenshots and High-Taste Motion Graphics

Jake plays the custom intro reel he generated for Peter's video. Peter praises the inclusion of an upfront table of contents:

> "I like listing the use cases because then people actually want to watch you see all 10."

Jake explains that the layout came from giving Claude reference screenshots rather than coding animations by hand. Peter had previously shared two video styles he admired. Jake took screenshots of those hooks, fed them into Claude, described how the apps and ideas should be showcased, and instructed the agent to skin the visual assets using Peter's signature channel palette (the "Behind the Craft" brand colors).

Peter and Jake then discuss finding the right balance for on-screen motion graphics so visual clutter does not distract from the person talking. Jake notes that the frontier models naturally lean toward understated design:

> "Let the A-roll do the work, right? At the end of the day, you want to learn from the speaker."

- Intro reels for educational content work best when they present an explicit table of contents or numbered list to drive retention. - AI agents translate visual concepts reliably when prompted with screenshot references paired with explicit color schemes. - Models like Claude tend to stay conservative with animation density, helping avoid over-edited, hyperactive visuals unless explicitly pushed.

[26:40](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1600s) Social Feeds vs. Long-Form YouTube Editing and the Power of Energy Prompts

Peter brings up creator Paul Lipsky, who recently created a talking-head video covering a blog post for Muse Wonder. Lipsky merely sat in front of the lens and talked; he provided Claude with a link to the blog post along with raw assets, and the model handled the rest—generating text overlays, inserting relevant screen captures, and highlighting the specific interface elements that corresponded to what he was explaining. Peter reflects:

> "I think that's the dream of a content creator you can just yap to the camera And then Claude will add all the screen shares and everything for you."

Jake agrees and contrasts the editing demands of different video formats: - **TikTok and Instagram Reels (Short-Form):** Require constant motion, frequent cuts, and heavy on-screen assets to halt rapid thumb-scrolling. - **YouTube (Long-Form):** Relies heavily on the title and thumbnail to secure the initial click, after which viewers appreciate cleaner pacing that lets the speaker teach without visual bombardment.

Jake highlights that current frontier models—such as Claude, Astra, and Gemini—excel at inferring context from external URLs and raw assets. The secret to steering their editorial intuition lies in prompting for emotional pace:

> "Describing the energy is so valuable to the agent."

Before HyperFrames introduced its latest visual style tools, Jake had to manually slow down reference clips frame by frame and write detailed prompt descriptions to get the model to imitate a creator's hook. Supplying direct blog links, sample assets, and descriptive "energy terms" now bridges that gap in a fraction of the time.

[30:00](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1800s) Reverse-Engineering Video Styles and AI Split Screens

Copying a visual style or kinetic effect from another creator used to be a frustrating, labor-intensive chore because of the dozens of minute details that make it work—such as the exact way text animates onto the screen or how background removal isolates a subject so secondary graphics can slip behind them. Jake explains that social media algorithms on X and Instagram are currently pushing a clear convergence toward specific editing hooks and kinetic effects. To solve this, HyperFrames added native style-cloning support: a user can drop in a short five-second reference clip, prompt the model on what specific elements to tweak, and receive a newly generated scene that borrows the structural mechanics while completely transforming the branding, typography, and color palette. Jake points out that even a fast-paced hook usually boils down to roughly ten discreet visual actions, making reference-based generation far more reliable than starting from a blank page.

```mermaid flowchart LR accTitle: Reverse-Engineering Video Style Workflow accDescr: A process showing how reference footage is uploaded, reverse-engineered into code, and customized with new assets. subgraph Input ["1. Reference Input"] A["Upload 5s Reference Video"] end subgraph Deconstruct ["2. AI Analysis"] B["HyperFrames Reverse-Engineers Layout & Motion to Code"] end subgraph Transform ["3. Customization"] C["Apply Custom Brand Kit & Colors"] D["Inject AI-Generated B-Roll / Talking Head"] E["Add Dynamic Captions & Split-Screen"] end A --> B B --> C B --> D B --> E ```

For experienced video editors who have refined their visual identity over years, this approach serves as a bridge into AI-assisted workflows. A creator can feed their existing library of finished videos into the model to deconstruct their formats into reusable, referenceable code templates. Peter notes that this also pairs naturally with modern split-screen setups, where an AI can automatically generate ambient B-roll footage (such as an airplane flying overhead) on one side while syncing speech and captions on the other. Jake highlights that the entire production chain—from the visuals to the talking head itself—is becoming radically hands-off: “If you don't want to record, you don't have to. If you do want to record, do it.”

- Creators can use screenshots, previous video files, or shared HyperFrames project code as direct visual references. - Reference clips are deconstructed into underlying code, allowing full parameter adjustments to fonts, colors, and motion graphics. - AI split-screen generation eliminates manual B-roll searching by prompting generative video or directing an agent to retrieve stock footage.

[32:55](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1975s) Product Launch "Moments" in the HyperFrames Desktop App

Beyond short-form social clips, Jake highlights how HyperFrames streamlines structured product launch videos using an internal concept called "moments." Because high-converting launch videos generally adhere to a four-part structure—the hook, the proof, the turn, and the payoff—the HyperFrames desktop application hosts a curated catalog of pre-built moments demonstrating how other successful launches executed each stage. Rather than hunting through GitHub repositories or scrolling through old X threads to find specific animation assets, users can browse visually, pick the pacing and layout they want, and map their own brand kit onto the template.

Jake references an internal guide where he outlines the three primary ways to prompt an AI agent when repurposing these launch moments: taking the structural energy of a clip, having the agent make the template entirely their own, or chaining multiple distinct moments together to serve an entirely new narrative arc. Peter recalls the friction of earlier setups where he had to dig up Jake’s Claude avatar typing code from a raw GitHub repo just to get the agent to understand the visual context. Jake explains that making these assets visually browseable inside a desktop app was essential, arguing that because video is an innately visual medium, users need to see and assemble reference components directly rather than relying on abstract text files or hidden repositories.

[36:17](https://www.youtube.com/watch?v=MC1g2fHYC88&t=2177s) Frame Annotation Feedback, Video Blur, and Long-Form Chunking

Addressing the everyday reality of AI editing, Peter brings up the frustration of iteration: while models occasionally hit a "one-shot" generation, fine-tuning edits using chat messages is tedious. Previously, Jake found himself scrambling to pause previews, memorize timestamps, and frantically type descriptive notes into Claude Chat like "at this moment I said this" to point out mistakes. To solve this, HyperFrames added an in-frame visual canvas where users can pause, draw directly onto the video frame to highlight an element, type an edit note, and queue up multiple annotations before submitting them to the model in a single batch.

```mermaid flowchart TD accTitle: HyperFrames Visual Feedback Loop accDescr: The loop of pausing video, drawing on-frame corrections, queuing notes, and batch-submitting to the agent. A["Review Video Preview"] --> B["Pause at Specific Frame"] B --> C["Draw Annotation Directly on Screen"] C --> D["Add Contextual Text Note"] D --> E["Queue Feedback Items"] E --> F["Batch Submit to AI Agent / Opus"] F --> A ```

Peter shares two practical production tips from his own setup: 1. **Automated PII blurring**: He created a custom agent skill called "video blur" that scans frame-by-frame across software screen recordings to blur out personal identifiable information (PII), sensitive emails, and names. 2. **Chunking long-form content**: For videos exceeding 30 to 60 minutes, Peter breaks the timeline into distinct segmented chapters before editing. Asking an LLM to reprocess a full 60-minute file to modify a small tweak at the 30-minute mark wastes thousands of tokens; editing isolated clips and having the agent stitch the finished segments together is significantly cheaper and faster.

Jake agrees, emphasizing that creators with recurring formats should always save successful prompts as persistent skills: “Once you nail that, just create a skill or similar. And all that is, is asking at the end of this session, hey, let's remember exactly what the steps here are so that I can reference it later for when I'm creating more videos, especially if you're somebody that has a consistent format that you use.”

[39:46](https://www.youtube.com/watch?v=MC1g2fHYC88&t=2386s) Democratization of Video and Wrapping Up

Peter and Jake reflect on how rapidly multimodal AI tools are lowering the barrier to high-quality video production. While expert human video editors will still hold an edge on high-end creative nuance, generative workflows now give average creators and product builders the ability to produce compelling, highly polished assets without deep technical expertise. Peter argues that dynamic video remains the most potent medium for communicating ideas online, half-jokingly pointing out that AI-generated rap battles—like the viral AI feud between ChatGPT and rival models—have emerged as an unexpectedly effective storytelling format that hooks audiences instantly with generative music and dynamic lyrics.

Jake shares where creators can access the software and connect with him: - **Desktop Application**: HyperFrames Studio is downloadable at `hyperframes.dev`. It is currently available for macOS and Linux, with a Windows release fully built and pending store approval. - **Open-Source Code**: Users can explore their open-source tooling under the `hyperframes` repository. - **Social Media**: Jake can be reached on X at `@jake11moran`.

Peter concludes the session by noting how inspiring the live demos were, committing to incorporating the tool's visual annotation, moment libraries, and reverse-engineering techniques directly into his own upcoming content pipeline.

## Notable quotes > [13:37](https://www.youtube.com/watch?v=MC1g2fHYC88&t=817s) The biggest bummer I think for people still today with editing with AI is that you actually, you probably do still need your idea, right? — Jake Moran

> [08:40](https://www.youtube.com/watch?v=MC1g2fHYC88&t=520s) I can't understate how powerful a zoom in is at the beginning of the video. It just leads your eye somewhere and kind of wraps you into the world. — Jake Moran

> [25:00](https://www.youtube.com/watch?v=MC1g2fHYC88&t=1500s) Let the A-roll do the work, right? At the end of the day, you want to learn from the speaker. — Jake Moran

> [02:30](https://www.youtube.com/watch?v=MC1g2fHYC88&t=150s) Getting as much context into your prompt, into wherever you're asking things to be edited is the key to increasing quality and getting closer to what you're wanting. — Jake Moran

## Mentioned Jake Moran, Peter Yang, Paul Lipsky, Claude, Claude 3.5 Sonnet, Claude 3 Opus, HyperFrames, HyperFrames Studio, Parakeet, NVIDIA, HeyGen, Suno, Wispr Flow, OpenAI, Google Gemini, Astra, Codex, Muse Wonder, Slack, TikTok, YouTube Shorts, Instagram Reels, X

## Video description Jake is my go-to expert for creating and editing videos with AI. In our episode, he demoed how he prompts Claude and HyperFrames to make incredible videos for product launches, long-form tutorials, and shorts. We also covered the key terms that you need to learn to guide AI to make the right edits.

Jake and I talked about:

(00:00) Video editing 101 to get Claude to make what you want (06:14) Parakeet: A free local model for accurate cuts (08:40) Hook viewers with zooms, B-roll, and whip pans (13:37) Demo: A launch video built from three screenshots (18:41) Launch videos in 4 parts: Hook, context, payoff, brand (20:39) Claude added Jake’s AI avatar without being asked (23:28) Jake’s AI edit of my raw YouTube video (29:25) How to copy any video’s style using AI (36:17) Draw feedback on the frame instead of typing

Thanks to our sponsors:

Wispr Flow: 4x faster than typing with your voice https://ref.wisprflow.ai/peteryang

📌 Get the takeaways: https://creatoreconomy.so/p/how-to-make-amazing-videos-with-claude-and-hyperframes-jake-moran

Where to find Jake:

X: https://x.com/jake11moran HyperFrames: https://hyperframes.dev Parakeet: https://build.nvidia.com/nvidia/parakeet-tdt-0_6b-v2/modelcard

Subscribe to this channel - more interviews coming soon!

[Source](https://www.youtube.com/watch?v=MC1g2fHYC88)

中文简报

HyperFrames利用Claude通过截图构建完整视频 AI视频工具可自动化时间线组装、跳切和风格复刻，使人类制作工作完全转向概念把控、品牌规范和帧级批注。 Jake Moran概述了一套AI辅助视频剪辑的端到端框架，该框架将Claude等推理模型与HyperFrames及NVIDIA的Parakeet相结合，以实现毫秒级精度的粗剪。他演示了如何利用四段式叙事结构将静态截图转化为精致的动态图形发布视频，整合HeyGen合成虚拟形象，并自动适配品牌色调。对话还涵盖了从简短参考片段中克隆视觉节奏、利用帧内视觉绘图画布进行批量反馈，以及分割长篇内容以规避Token限制的高级技巧。 HyperFrames利用基于HTML的架构，使Claude等原生代码AI智能体能够以编程方式，从原始素材和静态截图中组装产品发布视频与粗剪片段。 (0:00) 亚秒级视频微调需要精确的音频对齐；运行NVIDIA本地1 GB大小的Parakeet语音转文本模型，能够提供将停顿修剪至250毫秒且不切断音节所需的毫秒级时间戳精度。 (6:14) 产品发布视频通过遵循四步节奏结构来实现最佳观众转化：吸睛点、铺垫背景、成果展示或行动号召，以及品牌强化。 (20:00) HyperFrames允许用户直接在单个视频帧上绘制反馈，简化了剪辑迭代流程，并将视觉修正与修改意见排入队列以便批量提交给Claude。 (36:17)

Reading summary

Jake Moran demonstrates that frontier LLMs and web-native tools can automate technical video production—from rough-cut assembly to motion graphics—provided creators feed the models precise timestamps, explicit visual references, and brand context. Rough-Cut Automation Requires Precise Timestamps and Explicit Constraints (2:30) Modern LLMs like Claude 3.5 Sonnet, OpenAI models, and Google Gemini can eliminate dead air, false starts, and filler words in a single compound pass, but their cut quality depends entirely on timestamp accuracy. Low-resolution transcription engines cause agents to misjudge audio boundaries, cutting into adjacent words. To prevent this, Moran relies on NVIDIA's locally run 1 GB Parakeet model, which he reports achieves 98% reliability for millisecond-level word boundaries. This level of precision enables creators to strip thinking pauses down to 0.25 seconds or less for fast-paced social formats. Visual Framing, Pacing Curves, and Audio Prompting Guardrails (4:45) When raw cuts create jump cuts, editors mask the stitches using punch-ins, alternating focal lengths, and camera push-ins to lead viewer attention. Dynamic motion graphics rely on easing curves that accelerate through the middle and soften at the edges rather than moving linearly. While LLMs can handle layout decisions and choose transitions like whip pans, Moran cautions that models cannot hear audio. To prevent background music generated via Claude 3 Opus or external libraries from overpowering dialogue, users must provide explicit volume instructions, such as capping music loudness at 10% of the primary voiceover track. Structuring Launch Promos and Bridging External Generative Tools (13:37) For quick-turnaround feature launches running 30 to 45 seconds, Moran follows a four-part framework: an immediate visual hook, context building, the payoff or call to action, and the brand reminder. To bypass tedious manual animations, HyperFrames uses an HTML-based architecture that lets AI code visual assets directly from static screenshots or high-level narrative prompts. The platform also connects to external generative tools like HeyGen for digital avatars and Suno for music via direct app integrations or Model Context Protocol (MCP) servers, enabling synthetic demonstrations without filming on camera. Style Reverse-Engineering, Visual Feedback, and Long-Form Workflows (30:00) Instead of prompting motion styles from scratch, creators can upload five-second reference videos for the model to deconstruct into underlying code, allowing immediate re-skinning with brand kits derived from websites or PDFs. For iteration, HyperFrames replaces text-only chat corrections with an in-frame visual canvas where users draw annotations directly onto paused frames and batch-submit them to the model. Peter Yang notes that for 30-to-60-minute recordings, editors should chunk timelines into isolated chapters to save token costs and use specialized agent skills, such as automated frame-by-frame PII blurring. While AI tooling dramatically lowers technical production hurdles, Moran emphasizes that creative direction, narrative taste, and the core concept remain entirely human responsibilities. HyperFrames is currently available as desktop software for macOS and Linux, with a Windows build awaiting app store approval.

精炼阅读摘要

Jake Moran与Peter Yang在对谈中指出，前沿多模态大语言模型结合基于HTML架构的视频工具（如HyperFrames），正将视频剪辑从繁琐的时间线手动操作重塑为以文本和视觉提示驱动的自动化工作流。创作者依然需要负责核心创意与审美把控，但AI已能胜任粗剪清理、视觉包装及全流程生成。 高精度转录支撑文本化粗剪与镜头衔接 (2:30) 长格式录制或口播视频的原始素材通常冗长凌乱。现代大语言模型能够基于带时间戳的逐字文本快速完成粗剪，但提示词必须具体。Jake Moran建议在提示中明确设定清理规则，例如剔除“um”和“uh”等口头禅、丢弃失败的重录尝试仅保留最后一遍，并将停顿压缩至0.25秒以内。为掩盖跳剪造成的画面突变，可通过提示词驱动轻微推焦（punch-in）以制造有节奏的景别切换。 这套自动化流程的关键瓶颈在于时间戳精度。如果底层语音转录发生漂移，模型剪辑时就会切碎单词字头或字尾。Moran日常依赖体积约1GB的本地语音识别模型NVIDIA Parakeet，据其测试，该模型在提供词级精准时间戳上的准确率达到98%，能有效支撑TikTok、Reels等短视频所需的毫秒级硬切与快节奏压缩，同时避免外部API调用成本与延迟。 基于视觉参考与叙事框架构建爆款开篇 (8:40) 视频前几秒的钩子（hook）是决定留存率的核心。相较于线性移动，使用中间加速、两端平滑的缓动曲线（easing curves）配合缓慢推镜（push-in）或瞬间切入（cut-in），能够显著增强视觉代入感。Moran在产品发布视频中总结了一套四段式结构：最吸引人的钩子、使用场景与上下文构建、最终交付成果或行动号召（CTA），以及品牌信息回顾。在教育或深度访谈类长视频中，开篇列出明确的目录卡片也能大幅提高长线观看率。 在生成视觉资产时，直接向Claude输入参考截图并指定色彩体系，比手写复杂的动效代码更高效。此外，AI模型无法直接“听”音频，在混音时往往难以把握平衡。Moran指出，创作者必须给出明确的量化提示（如强制要求背景音乐音量仅占主音轨语音的10%），同时可要求AI将剪辑切点精准对齐音乐节拍以提升成片质感。 从三张截图到全虚拟资产的快速成片流 (13:37) 在初创团队快节奏发布功能的场景下，传统制作流程难以匹配交付速度。Moran演示了如何仅凭Studio界面截图与两张视觉参考，通过向Claude输入高阶叙事提示，在HyperFrames内生成背景消除功能的发布视频。为节省反复完整渲染的计算成本，该流程先生成静帧分镜模式（storyboard mode）供创作者微调构图、文字排版与节奏，确认无误后再进行全动态渲染。 AI驱动的工作流甚至打破了真人出镜的限制。在一次演示中，画面包含由HeyGen生成的庞大数字人网格墙，Claude甚至在未经明确指令的情况下自主调用并插入了Moran本人的AI数字分身。通过HyperFrames的原生模型接口，或在Claude Desktop与Codex中利用CLI/MCP协议挂接HeyGen、Suno等外部生成工具，用户可以全自动组装视听元素，摆脱传统外景采集的束缚。 风格逆向工程与带标记的批处理反馈机制 (30:00) 社交媒体上的视觉风格往往高度趋同。HyperFrames支持上传约5秒的参考片段，AI会自动将其拆解为底层代码结构，随后允许创作者套入自身品牌色彩、字体及AI分屏B-roll，实现低成本风格克隆。为了避免AI在品牌风格上出现幻觉，系统能够自动从公司官网或上传的PDF规范中抓取调色板并转化为品牌资产包。 针对AI编辑微调困难的问题，传统的暂停记秒数并在对话框描述修改意见的方式极易出错。HyperFrames引入了画框视觉批注画布（in-frame visual canvas），创作者可直接在画面上暂停、框选涂鸦并附带文字说明，打包后批量提交给模型。针对超长视频，Peter Yang分享了工程实践：将30至60分钟的内容切分为章节独立送审以节省Token，并利用自定义AI技能自动逐帧扫描并模糊屏幕中的个人隐私数据（PII）。 当前多模态AI大幅降低了视频制作的技术壁垒，使独立创作者能够快速产出媲美专业水准的宣发与口播内容。但Moran强调，AI无法替代核心创意与审美判断；且该工作流高度依赖底层转录精度与结构化上下文输入，提示词的精确度直接决定了成片质量。

Source quotes

[13:37] Jake Moran: The biggest bummer I think for people still today with editing with AI is that you actually, you probably do still need your idea, right? [8:40] Jake Moran: I can't understate how powerful a zoom in is at the beginning of the video. It just leads your eye somewhere and kind of wraps you into the world. [25:00] Jake Moran: Let the A-roll do the work, right? At the end of the day, you want to learn from the speaker. [2:30] Jake Moran: Getting as much context into your prompt, into wherever you're asking things to be edited is the key to increasing quality and getting closer to what you're wanting.

Original transcript

[0:00] It came with three screenshots. It came up with so much of this on its own. None of these people are real except for you in the middle there. Or even you is not real. That's my avatar as well. Yeah, and I didn't even ask for myself to be in there. That was actually Claude's discretion. You just upload a reference. So I uploaded this five seconds and then I described in the prompt just what I wanted to change. You can just draw on what you want to change, leave the note and then they'll queue. screen shares and everything for you. All right, hey everyone. Today I'm really excited to welcome back Jake from Hyperframes. Jake has only been editing videos for six months, but to me, he's the go-to expert on how to use AI to edit videos. Today we're going to walk through video editing one-on-one and also how to edit product launch videos, talking head YouTube videos, and also short form videos. So welcome, Jake. Hey, Peter. Thanks so much for having me. Why don't we start with this? I'm like a total new when it comes to video editing. Can you walk us through some video editing concepts and terms and what they might look like so that we can all get educated? Video creation, it's a super wide net. So there's tons of different types. There's launch videos, which is what I've worked on primarily for the last few months. I'm actually just now really getting into the weeds of editing something like this, which is a podcast where we've recorded this and then we need to do some kind of cleanup later. And I want to try and visualize a couple of these things here today. When we finish recording this video, we're going to have something called the or the raw takes, which is going to be maybe slightly longer than what the finished video will look like. And so in the case of using AI, it makes it a lot easier to... get to this kind of shortened version that's a little bit more polished. There's a couple terms that are worth knowing, and I'm sure Peter's gonna be able to share this out in some form. later so that we have a visual reference as well. For me, I use Hyperframes as my editor. There's plenty of other editors out there. Hyperframes is HTML based, so it's a pretty native way for agents to edit. So I find a lot of these techniques to work pretty much one shot if you ask for them. So there's in terms of like cleaning up recordings before you put anything crazy on the screen, there's something called the rough cut, which you can ask your agent for. This is basically where we take out a lot of the mistakes that might have happened during recording and get back to that consistent storyline that you wanted when you... started recording. The way that I would do a rough cut is I would say, let's remove things like filler words, which are all the things that I'll probably say right now on this podcast, I would normally want to remove to get to the meat of what I'm trying to talk about. The same thing, I'm probably going to stop and think about something before I say it, which would be Thank you. language that both you and the LLM will recognize because that'll help you get to results faster, not waste tokens. We all know that AI is expensive. Editing definitely isn't outside of that. So there's different things that you can ask to be removed because models can do transcription, which is basically taking the audio and understanding what was spoken from that. So It can find filler words. It can find the pauses because it'll take timestamps. It can find moments where I say something twice. It can also find things where I might start saying something and then I'll realize that I might have blundered and I say it again. All of these can be found. The key is that if you really want to maybe guarantee that you're going to get closer to that cut down that you want that first go is if you use these types of terms, which is, "Hey, I think I had a lot of filler words. Here's an example." I might have said, "I want to remove ums." I think there was a moment where I was Pausing a lot, I would like to cut all the pauses to around... let's say a quarter of a second or less. And it goes down the list in the same way, right? I might have been repeating words, I want to get rid of those. I know that I did a retake around here. Basically, I think this is true with all things AI, but it's also especially true with video editing. Getting as much context into your prompt, into wherever you're asking things to be edited is the key to increasing quality and getting closer to what you're wanting. So the biggest thing you can do is try and explain as much as possible. And this, just to finish it, the jump cut and the punching. So jump cut is like all of a sudden you jump cut to saying something else and the punching tries to hide some of that stuff, right? [5:00] Yeah, so same kind of thing. If I The way the LLM will probably do it is if I say if two times, it'll cut out one of those ifs and then move to whatever the next thing is. item is. And so that will just... to the viewer, they might notice it just because it will appear as a slight, oh, my head might have moved from here to here because in between that, the time has been taken out. And then the same thing here, if you want to move something that you said later to earlier in the video, that would be the same idea where you're punching in this clip that would add additional context for the viewer at the right time in the video. So these are all ways to use your footage that might have happened thing. Got it. And basically you have individual prompts here, but usually what I do is I just combine all the stuff into one prompt, right? Just like remove all the filler words, repeat phrases, just jump cuts, use punchings. Yeah. Yep. Yeah. And this only happened really in the past, basically with Opus 5.5 and maybe with Astra and Gemini 4. So really only the past few months, right? It was able to kind of get to this level. And for this, for the main video, I had to look at the timeline and look at the transcript and cut stuff out. It just took forever, man. So I think this is a massive time saver. And you also had a tip that you told me because some of the transcription models aren't as good. Can you share that tip? I think you use Parakeet or something. Yeah. So I use Parakeet, which is a locally hosted model. I think it's about a gigabyte in size. So it's big as local models are, but it's worth it. So basically all of these tools, no matter what you use, either if they're operated from MCP, which would be the kind of older generation tools. That's how AI interacts with them or us. I normally have some kind of transcription software. For Hyperframes, we have a smaller size local transcription model. And the thing is, sometimes it misses on the exact timestamp. It's a bit hard for LLMs to figure out exactly when something's set down to the millisecond, but it's key for, let's say Instagram or TikTok based editing, where it really, really matters that you're super concise and that the pacing is really high. And so you might ask to be like, "Okay, cut down the pauses in between sentences to a quarter of a second. with a bad transcription model, the LLM will think, okay, cool. And then it might be operating on some bad timestamps. And then you'll have like half your word cut off or things like that, that are awkward. So Parakeet for me works 98% of the time to getting the right amount of time in between words or whatever the case you need. It just allows the LLM to get really good timestamps, which is everything for these types of, let's say, recording cleanup. Awesome, yeah, we can link the pair key model. It's free in the video description. This episode is brought to you by Whisperflow. Whisperflow saves me at least three hours a week and is one of my favorite AI apps by far. It's just so much faster to dictate to AI using your voice than to type. You just talk naturally and it outputs clean, ready-to-send text. Whisperflow even removes filler words and formats your sentences for you. I use Whisperflow for everything including drafting newsletter posts, writing product specs, replying on Slack, and more. It at whisperflow.com and use my code PeterWhisperflow to get six months free. That's PeterWhisperflow. Now back to our episode. - So that's the main video, like getting rid of the repeat phrases and all that kind of crap. What about the intro? The intro's trying to hook the viewer, right? I think you have some more stuff down here. - Yeah, so intro, hook are normally the main terms you're going to hear. I think the biggest thing is if you're trying to get into this game of creating video, you're going to hear all of these buzzwords. And a lot of them mean the same thing. Intro and hook are generally the same thing. And normally what you want to do in these, if you're editing existing footage, I'll go into it with motion graphics as well. But you'd want... and then you can see that the image is a little bit more engaging. So, if you're editing existing footage, you can see that the image is a little bit more engaging. I can't understate how powerful a zoom in is at the beginning of the video. It just leads your eye somewhere and kind of wraps you into the world. And there's tons of different ways to do that. You can ask an LLM right now to just, you know, just go to the world and you can just go to the world and you can just go to the world. how to zoom in in the beginning and it'll probably look pretty good. There's different types of zoom in. So there's push ins where you start further away and then the camera kind of appears to come in closer like this and then there's There's cut-in where we don't actually see the zooming happen, [10:00] It just jumps from one position frame to another. And all of these are controlled by a certain speed. Most creators you'll see are using a curve that looks something like the one on the right here, where we start slow, kind of ramp up, and then smooth out towards the end. Then the next types of things you can add to a intro or hook is B-roll. So this right now that we're hovering on is all A-Roll, but this would be an example of B-Roll, which is something that you might be talking about later in the video. It might be something that's just for the purpose of being visually engaging. It might be maybe a table of contents for a more tutorial style thing. If you're somebody like Peter that has a big audience, you might put your overlay explaining, hey, here's the reason you listen to me because I've got a big audience. I've got a long history, right? Like there's certain pedigree that's normally used for these speaker cards, things like that. So these are a couple examples of other types of techniques that people use. So transition in this case is pretty broad, but the one that's shown here is just a cut where there's footage to scene B, we just straight drop in. The second one on the right is called a whip pan, where it appears that the scene A would go out to the left, and then scene B enters to the left. You can do this in any direction. It's just another example of more visually engaging ways to hook in viewers. Nice. These are examples as well. Other kinds of transitions. And maybe you can just like, maybe you can just like, because I don't actually know when to with pan versus cut. So maybe, maybe you just like mention all the transition names or like a bunch of them and be like, hey, cloud, use your best judgment. to pick it up. Exactly. Honestly, yeah, it is rare that I'm deciding what exactly needs to happen at when. It's more, hey, I like these effects, right? When I see another creator use, let's say, the wet pan, I like it. I think visually, it's just a really engaging type of effect. Leave it up to Claude, it's going to make the choice for you, and it'll probably be pretty solid. Got it. Awesome, dude. And then you can also ask it to add music. Do you have to use or you can just randomly add music by itself? - Yeah, so this one's super new to me, and I'm still experimenting with it, but Opus can make its own music too. So it can use MIDI and create its own tracks. You can point it at different types of libraries to create its own music. If you just ask, hey, don't use a generative model, make your own music. There's, you can, if you have local files on your desktop, which I do, I just drop it in chat. this music. One thing that kind of goes back to the idea with transcription as well. Models still struggle with audio. They can't hear anything. And so often if you just say, put the music in, it might come back super loud against what you're talking about and against your voiceover. So I normally just say, keep it low, keep it at 10% of the volume of my A-roll. There's all sorts of ways to describe this, but just being a little bit clear normally helps. If you have an idea that something might happen or you have an idea of what you want, it's better to just list it all out before sending the agent off on some kind of list. task. Awesome. Yeah, this is awesome. And of course, asking it to cut the transitions to the beat of the music will take it to another level, right? That's kind of... Exactly. Yeah. Dude, so do you want to show us a finished video that you made? Should we start the product launch one, or should we start with something else? So this is something we launched last week, which is a feature inside of our desktop app, Hyperframes desktop app that allows people to remove background on video and images. What I'm really going to actually be going over is more of "Okay, how did I actually make this video?" Right? There's the meta idea of we have this feature, how do we describe it to people? And so I came up with this. Thank you. All right, that's pretty great. That's pretty good. So yeah, let's see how you made this. Yeah, so let me open up. In our case, I made it in our desktop app. You can also use [15:00] I'm just going to show our app because that's where I made it. But the creation flow is largely the same thing. This just made it easier for me to edit later. Basically, I came with three screenshots. I'm just going to show our app because that's where I made it. That's what I came with when I wanted to start this video. So one screenshot I had was of our actual studio because I knew I wanted to share it later. I came with the images I wanted to use in Hook and then I sat down and I wrote this prompt. Kind of core to what I was going over earlier, it's the same idea with motion graphics videos as well. As much context as you can feed is ideal. But this isn't It isn't that much. And the key is that I'm not actually using really much editing jargon at all, which goes back to that idea that, like, Opus and these newer models, they're gonna make a lot of these choices themselves, and you'll be amazed once you get started. by what they choose to do by themselves. Really going over in this is the storyline. that I wanted. And I will say, and this has come up before, the biggest bummer I think for people still today with editing with AIs that you actually, you probably do still need your idea, right? You probably still need to know a rough storyline that you want to convey. And so in the case of this video, when we were creating the background removal feature, one of the engineers on my team had actually shared this hook sequence basically with me in Slack and was like, "Look, we have background removal and it works," and showed this image with the background removed. And I was like, "Okay, "That was super cool, I wanna use it in the video." So I came to create my launch video with the same asset that he had shared, dropped it in, and then loosely in the chat, described what I want. which is literally not that much information, right? It came up with so much of this on its own. And then the rest of it, I just wanted to explain kind of the, now it's just a prompt away, like the idea of just normal, how people do the more marketing-y type things in these videos where it's just like, oh, it's a prompt away, right? Things like that. But it wasn't like, okay, now we're going to cut into this and it's going to be 35 seconds and all of the very specific things a prompt like that anymore. It's more about get out what you're thinking, and you'll get something back. And the last thing is I normally ask for a storyboard, in this case, in this tool, we actually have this as a thing you can select, so I had that selected. And so, The LOM goes off. and it makes something like this if you ask for it. Otherwise, it'll probably go build the full video, which takes a long time. I prefer, just because motion graphics, I have a certain standard I want to get to, I prefer to look at still frames of what's going to happen over the video instead of kind of working on the fully built thing. But basically, going back to that full video that we saw, you can see all of the pieces that actually ended up making it to the finished film, but this allowed me to edit things first. So you can see I didn't end up having this exact look, so I asked for more things later when I got the first build back. I normally just use this as a rule of thumb, where I don't need it to be perfect, but if I can kind of see what I want, then I'm good to let it build. but it allows you to catch some of the things that might be issues a lot faster. And generally speaking for product launch videos, I think you've made the most of these, right? So just as a high level, do you have any quick guidelines for making these? Like, you know, they're usually like 30 to 45 seconds long. They just tell a story and maybe like bring the problem or show the solution. - Yeah, in our case, mainly what we're doing with product launch videos is it's some feature, which is a subset of our larger product. I'm not often doing like a... Like, here's this new company, I have to explain the founder story and all the things that you might see for when people announce a round or something like that. Often what people really need is, hey, I just made this tool. In our case, our engineers go crazy all the time. So I'm getting pinged at 5pm being like, hey, we just finished this. Do you think we could talk about it tomorrow? Right? So it's some smaller scope item that is useful to something or some subset of users. concise They don't need to know a ton of things. So you try and get somewhere, or generally I try and get somewhere in like 30 to a minute and change, 30 seconds to a minute and change, and I start with some kind of engaging hook. In this case, it was that. basketball thing, but in general, [20:00] And I'll get into different ways to add context in just a second, but the hook is your most engaging thing. Then you kind of want to lead to how to use it in the product generally, and then what the payoff is going to be for the end user. And then maybe you need some kind of follow-up information about if it costs money or something like that, or just like, Hey, we're hyper friends again, by the way, right? Just reminding people of the brand. So I normally think of that as a, like, I'm just going to get into the product. normally think about it in four sections, hook, kind of like the context building, maybe the call to action or the output, and then coming back around to like, hey, here's our brand. Got it. And a couple of quick questions. So I think the avatars are actually not generated by cloud. You're using Heijen, right? To make the people? Yeah. So I don't like being on camera. I also don't have access to a ton of creators. So I, being a HeyGen employee, being someone that's been around HeyGen way before I started doing this Hyperframe stuff, I've gotten super comfortable with just using avatars to generate assets that I need in videos In this case like we see later in this video right we have this crazy grid wall of video examples I could have pulled these from online I lazy And I just asked Claude to use Heyjen to generate different characters and then have them speak a couple clips. In this way, I could get the idea that I wanted to get across more easily and without me doing any kind of research work on Instagram and TikTok and pulling videos and all the things. So none of these people are real except for you in the middle there. Is that right? Exactly. Or even you is not real, right? Is it like an avatar? No, yeah. That's my avatar as well. Yeah, and I didn't even ask for myself to be in there. That was actually Claude's discretion. If you have footage, you can use it. You would just drop it in. In this case, I really didn't want to source it. I have access to these models. Why not use them? It sped up my workflow and made mine easier. And you didn't actually go to heygen.com or whatever, right? You just used the MCP and just talked to Claude to do it? Yeah, so in our desktop app, we actually have a direct connection to generative models. In the case you were using it in the Cloud desktop app or let's say Codex or anything like that, you would have to connect via CLI or MCP. And that's true for any of these other generative features. You can use ImageGen audio generation like Suno All of these require some kind of outside connector to generate Okay And the other question I had was how do you make sure that the brand is consistent across these different videos? Do you have a skill with color and font and stuff like that? Yeah. So we have a feature now that allows you to create a brand kit just from your website. In this case, I already had brand kits. Let's say the HeyGen one where I can edit all of these items myself, but the agent just pulls them down from your website. Or you can upload them. Let's say you have a PDF with all of the information or a folder. Got it. But this helps so much. You do need to add it in. If you just say my brand, probably nine times out of 10, you're going to get some crazy colors and an imagined surface. So it does matter to add the context. For us, we built this natively. You just have to be more specific if you use it outside of our app. Okay, yeah, this is awesome. Okay, awesome. Yeah, so that's product launch video. And why don't you show us the other ones now? I think you actually edited one of my videos, right? One of my talking head stuff. Yeah I did Basically I might have gotten access to a raw unedited version of one of your episodes And now basically I did And this was 20 minutes and 30 seconds long. And then when I finished editing it, it was 18 minutes and 11 seconds. And that comes back to the thing that we were talking about earlier, where there were items that were maybe retakes or pauses or things that in the long run weren't necessarily needed. So the finished output ends up being shorter than the raw footage that comes in. And that came from that initial cut down of the video that generally, I mean, who knows how much will be cut out? Maybe I've recorded for eight minutes before and because I was sloppy and didn't have a script and whatever, the finished output of what I actually needed was only a minute, right? That's that you never really know it just depends on how let's say clean you were through the creation process there But there are two things I really wanted to do in this example and we're not going to play the full 18 minute thing But I wanted to work on a fun intro reel that I could show off for this episode And then just going back to this stuff I wanted to cut out all the things that I might have thought have been [25:00] Pauses, repeated words, retakes, et cetera. So I ended up creating this intro reel. Thank you. All right. *sigh* Nice. I like listing the use cases because then people actually want to watch you see all 10. So that makes sense. Exactly. So that was and and I didn't I didn't choose what should be the the. where that cut-in should be or anything like that, but it goes back to that idea of if we're going to do a tutorial, most likely you're going to have some kind of table of contents and you're going to introduce that to try and lead people into this video. And so for these effects, I had two examples that Peter actually shared with me that he liked. And all I did was shared screenshots to the agent of what I was going after. And then I said, I think I want to do in the case of that, that first thing where we showed the apps, I just described that, right? So I had a screenshot of somebody else doing a similar type of hook, and then in this case, it was just about translating that idea. to the contents of this video and then also switching it to the behind the craft colors that you use for most of your videos. Nice. I think one thing with this intro stuff is you want to add some animations and motion graphics, but you don't want to overwhelm with the actual person speaking. So you have to do it in the high taste way, right? You don't want to just add crazy animations all over the place. Yeah. I mean, some people do. I think I definitely agree with you. Let the A-roll do the work, right? At the end of the day, you want to learn from the speaker. We don't need to, like, I don't think we need to go crazy. Some people prefer that look, a long way and the models seem to be pretty good about this they seem to be maybe more conservative than i normally am so yeah okay so basically to make this intro real you basically just like submitted like some sort of prompt of like hey you know at the table of contents at the floor diagram and so on right exactly okay awesome you know what i think just real quick on this talking head stuff showing paul lipsky here i actually just talked to him and it's actually kind of incredible how far this thing has advanced he made a video i think about muse wonder that uh into the camera all the other screen sharing and effects were made by claude so so basically he he just kind of shared a link to the muse wonder blog post and maybe some other assets and then claude made all the text effects and i highlighted the right part where he's speaking and i haven't tried this myself but it's pretty incredible because i think that's the dream of a content creator you can just yap to the camera And then Claude will add all the screen shares and everything for you. Yes. No, yeah, exactly. And I think there is a fundamental difference between, let's say a YouTube video versus these more short form that end up on TikTok and Instagram. Normally you, you lean a little bit more towards the, a lot of effects or more busy on social media, just because we know it's harder to stop people from just scrolling on to the next video. Whereas YouTube, The commitment normally comes more from the thumbnail and the other things that get people to click in. For those types of videos, again, Claude also now and Astra and Gemini and all of these kind of leading frontier models, they all also seem to have this sense, it's so crazy. You can probably just give the link to the blog post, you can give maybe a couple other assets if you have them on your computer and a rough description. Describing the energy is so valuable to the agent. If you're just like, I want this to be I like to say energy type terms work really well for helping these models get there. We added one feature that I'll show just for another way of doing it. Basically, it all comes back to how can I add more context to the agent? Giving a link to a blog post really helps. The other thing is, you know, you can get a link to a blog post. If you have styles that you really want to emulate, You can either watch them really slowly and try and describe them, which let's say I really liked this creator's effect for a hook or an intro for social media. How I was creating videos in the last four months, five months before the last... couple weeks when we unlocked this new tool internally and we shared it yesterday. If I liked this effect, I would just describe it, right? Which... [30:00] Saying that just... if you really want to copy someone's effect and then turn it into your own, it's not the simplest thing. As you can imagine, there's a lot of little nuances in what just happened in this hook, including how the words enter, how there's background removal so that some of the items can be slightly behind the character. There's a lot that needs to be... described and that's what I would kind of struggle with in previous months. Now, if you scroll X or if you scroll Instagram or whatever, if your algorithm is remotely involved in this game, you'll see people are learning how to recreate certain effects in different ways. Yesterday, we added this as a native capability to, if you just upload a reference, so I uploaded this five seconds, and then I described in the prompt just what I wanted to change, right? Because you can see where we're drawing inspiration from, But... at the end of the day, this feels like a very different scene because there's different font, different colors. But... It's still structurally the same idea. And a lot of people, when you... So if you look on social media or video sharing platforms, there is kind of a convergence, I would say, happening on certain effects. And a lot of people want to use those effects. So this gives you the ability to create those or you can just Try and discard them, right? There were only probably 10 total things that happened in this hook. It's more about getting those right. It's easier if you can start with references. May that be screenshots. May that be previous videos. May that be somebody else's hyperframes project where you have the code to use. All of these are valid paths to getting to what you want. So basically, the way this is, first you uploaded the clip. and then you asked AI to reverse engineer it, and then you added your own tweaks to it, right? - Exactly, just for what I wanted in my video. And you can do this with all different types of things. So this is like a split-screen effect, the motion graphics on top were generated by hyperframes, we added captions. The other thing too is existing creators, right? If you're someone that's been editing for a long time, maybe, and you're looking to get into AI editing, this would be really useful because you just want a one-to-one of what you've probably already gotten down to. You have your style, you have your colors, you have your formats. All of these things are dialed down through years of creation. You now have a library that can be referenceable if you reverse engineer it into code. And so that's another path for the more already creating type of people as well. And I like you're showing the split screen view because now you can actually use AI to generate the B-roll footage on the split screen. Like probably that playing is probably just AI generated. Yeah, because before you had to go find all the bureaus and customize it. It will take forever. But now you can just like generate the stuff. And in like the agent can also probably find real footage to download it and then use it as well. So all of that, a lot of it, it's just becoming more and more hands off and more and more focusing on what you're trying to convey exactly in that moment. And that's what's making it more fun for me as well, trying to ramp up how much we can share about... tools like this online. It's getting easier for even myself to talk about creating with these tools because they're getting so much better. I think honestly like the arrow, the person talking can probably just be AI generated too. I mean, yeah, I mean, it's up to you. Yeah. If you don't want to record, you don't have to. If you do want to record, Do it. It's all kind of the same process at the end of the day. It's just more up to your discretion. It's great. Awesome. Well, anything else real quick you want to show on the short side or is pretty much? I do want to show one thing just for product launch videos, because I think maybe if people see hyperframes, they might want a little bit of additional information. info here. So we have these things in our desktop app. that we call moments. That's an internal term, so I don't I don't believe... Everyone will know that right away, but it's going back to that idea of there's probably four main sections in what you create in a launch video. So like the hook, the proof, the turn. These are all just ways of expressing sections. And what we have here is examples of how people have done this before. And All of these are available to anyone if they use our desktop app. And all you have to do is add them and then add a little bit of additional information. And I have this article here where I visually just quickly show the three ways that we use Inspirations internally. And so there are three ways that I would ask the agent to use a section for something that I'm creating, which is either make it Ours, if I'm using a bunch of them, [35:00] - I would say something like this, each one doing its own thing, serving the purpose of the new story or I want to take from some of the energy. And you can see these are the actual prompts that went in, which is basically this is the same idea, just the context that we have lunch, and then it's it's that same get inspired by... Got it. So it's really lightweight. You add the moments that you like and then it'll recreate within your brand context because my brand kit was selected. Yeah, this makes it easier instead of... Because I remember you had like a cloud video of the cloud avatar. typing stuff and before I had to go find it in the GitHub, right? Somewhere. Exactly. Before my agent would have to find it. But now I can't actually find it. It just, at the end of the day, for us, the desktop app allows us to visualize a lot of the things that we've been, that have been hidden in some repo or whatever it is, like some X article or something like that. It just allows, because video creation is so visual for, I mean, it is, it's so much easier as the end user to have everything available to you to see and then reference when you're ready to start creating. So I just wanted to show that I'll also link this article I'm sure with. Can you also show one more thing? Okay. Let's be real here. Many times this thing can one-shot it, but a lot of other times you have to give it feedback. And I honestly haven't tried Hyperfearing Studios before. I'll probably try it right after this. But when I just use Cloud, I have to give it feedback. I'll be like, hey, five minutes in, go do this. Ten minutes in, go do that. And it's kind of annoying to type that stuff out in text. There probably is a way to annotate this stuff, right? In the frame. two weeks ago when we had this desktop app internally, edited of the same format, which was I watched this whole thing and then I'm trying to frantically leave notes as I go. Like I'm watching it and typing in Claw Chat at the same time. And that's difficult, right? I'm trying to like maybe call back and be like, well, at this moment I said this to give some kind of additional reference or I do the timestamp if I catch it. But at the end of the day, it, It just was frustrating because, yes, there are going to be edits. Maybe Opus doesn't get it right. Or after watching it, you have a better idea. Half the time I'm creating, I'm like, oh, wait, no, this would be sick. I should come up with some other thing. and that's just part of the creation process. For us, we added this ability to, let's say, I wanted to change something right here. you can just draw on what you want to change, leave the note, and then they'll queue. Let's change. Let's change the text. Oh my god. So It'll queue here. And then when I'm ready and I've watched through the whole video, Um... Let's see. Instagram. I can just send these off, which just feels like to me, it's so much easier to edit using that type of experience than trying to rattle things off. I still sometimes do it if I have like a 10 second video, but yeah, it just makes it a lot more comfortable as the viewer to make edits. yeah and i guess a few more tips from my side so one tip is i've built a skill called video blur that blurs out all the sensitive information because i'm usually trying to screen shares like clock code or something and there's some pi stuff in there yeah and then definitely so first i get to do a video blur pass and it just goes frame by frame and blurs out like emails and names and stuff that's super helpful and i think also with longer videos anything longer than let's I like to split it into different segments. So it's also really good at stitching videos together. Because otherwise, if you want to make a change at like 30 minutes in to an hour-long video, it has to reprocess the whole thing. We're just going to like just burn a bunch of tokens. So I actually kind of like spin into different segments and edit one by one and then stitch it together or get AI to do it. Yes, yeah, 100%. Yeah. We're in a funny time where there's all sorts of things like this, have an idea of how it would be faster to be done with AI. Once you nail that, just create a skill or similar. And all that is, is asking at the end of this session, hey, let's remember exactly what the steps here are so that I can reference it later for when I'm creating more videos, especially if you're somebody that has a consistent format that you use. You definitely want to save things as. a reference so you don't have to do the same work twice. Yeah. All right, man. Well, it's crazy times that we live in. I guess this really democratizes video editing. I think if you're a professional video editor, you can probably still do a better job, but at least the rest of us can now get decent enough job. And I think [40:00] really powerful because i think video is actually the most effective way to communicate something right instead of using slides or text for a lot of things so yeah i i highly encourage everyone watching this to get opus to make launch videos for everything or help it edit your videos i feel like the ultimate format is a rap battle for everything like like a rap battle video you know then then people would just love watching it so didn't you didn't you make a rap battle video from ChatGPT and I don't think they're actually feuding it anymore but yeah it was pretty amazing. Yeah, there's so many more things also with AI music and all the things like that that you can just create kind of whatever you're thinking about and be surprised by what these frontier models are able to do from probably not that much information, right? I can't imagine that that was superscripted. Yeah, it is extremely fun. Yeah. All right, dude. Well, so where can people find you and Hyperframe Studio Studio? Yeah. So Hyperframe Studio is available to download on hyperframes.dev. Right now it's only for Mac and Linux. We are trying so hard to get Windows out. We're hearing everyone talk about it. Hopefully sometime within the next few days, we're waiting on their side. It's not that the app is built. So we're just waiting to be approved there. Otherwise, you can find our repo, which is also called Hyperframes. agent for me i'm on x you can find me at jake 11 moran m-o-r-i-n awesome awesome yeah thanks for bringing this tool and showing us uh your pro pro tips i'm really inspiring i'm gonna definitely do a lot of stuff myself thank you so much