Peter Yang Cuts 90% of Podcast Production Work by Chaining 8 ChatGPT Skills

Peter Yang

Having AI diff its own drafts against your manual edits and bake the changes back into its prompt instructions is what trains modular skills to eliminate repetitive revisions without producing low-quality slop.

AI-assisted summary of My 3-Step System to Automate Content (ChatGPT + Riverside) — original source. Check the original for full context.

Source published: 2026-09-16

Key ideas

Peter Yang developed an interconnected system of custom ChatGPT skills and MCP integrations that automates roughly 90% of his 15-step weekly podcast workflow, saving at least five hours per episode. (0:00)
Connecting external services via Model Context Protocol (MCP)—such as Riverside for automated audio/video trimming and Linear for editor ticketing—allows conversational AI to directly manipulate media files and assign project tasks. (6:42)
To systematically eliminate recurring draft revisions, Yang uses a 'snapshot diffing' feedback loop where ChatGPT compares its initial output against human edits and generates updated prompt rules to achieve one-shot accuracy over time. (13:09)
Content repurposing is coordinated through a master orchestration skill that triggers specialized sub-skills to generate Substack articles, YouTube packaging, show notes, and social clips routed into scheduling tools like Typefully. (9:31)

Detailed notes

## Background Published on YouTube by Peter Yang on September 16, 2026, this video details a three-step system combining ChatGPT and Riverside to reclaim at least five hours a week. According to the show description, Yang breaks down how he built eight modular AI skills to automate guest research, interview preparation, editor handoffs, Substack newsletters, YouTube packaging, and social clip distribution.

## TL;DR Yang breaks down his end-to-end podcast automation system by first mapping his 15-step manual workflow, creating dedicated AI skills for prep, editing, and post-production, and orchestrating them into an integrated pipeline. Walking through a real episode featuring OpenAI's Ethan, he demonstrates how ChatGPT skills connect with tools like Riverside MCP, Linear, browser automation, and Typefully to generate interview guides, editor briefs, thumbnails, show notes, and scheduled social clips. Crucially, he highlights human-in-the-loop oversight and introduces a snapshot-diffing technique to compare AI drafts against manual edits, continually training the skills toward one-shot accuracy while eliminating low-quality AI slop.

## Atmosphere & style The video carries a grounded, pragmatic, and refreshingly candid tone that cuts through Silicon Valley AI hype to focus on repeatable utility. Yang communicates with the methodical, calm efficiency of an experienced product manager, openly acknowledging where first-pass AI drafts fall short and showing the unglamorous editorial work required to tune them. It feels like an authentic over-the-shoulder work session—unpretentious, instructional, and deeply committed to maintaining high standards of craft even when delegating heavily to automated systems.

## Bulletin - [00:00](https://www.youtube.com/watch?v=PZHBW6yld5U&t=0s) Peter Yang developed an interconnected system of custom ChatGPT skills and MCP integrations that automates roughly 90% of his 15-step weekly podcast workflow, saving at least five hours per episode. - [06:42](https://www.youtube.com/watch?v=PZHBW6yld5U&t=402s) Connecting external services via Model Context Protocol (MCP)—such as Riverside for automated audio/video trimming and Linear for editor ticketing—allows conversational AI to directly manipulate media files and assign project tasks. - [13:09](https://www.youtube.com/watch?v=PZHBW6yld5U&t=789s) To systematically eliminate recurring draft revisions, Yang uses a 'snapshot diffing' feedback loop where ChatGPT compares its initial output against human edits and generates updated prompt rules to achieve one-shot accuracy over time. - [09:31](https://www.youtube.com/watch?v=PZHBW6yld5U&t=571s) Content repurposing is coordinated through a master orchestration skill that triggers specialized sub-skills to generate Substack articles, YouTube packaging, show notes, and social clips routed into scheduling tools like Typefully.

## Key takeaways - [01:34](https://www.youtube.com/watch?v=PZHBW6yld5U&t=94s) Documenting every manual step across prep, editing, and post-production is an essential prerequisite before designing functional AI automation skills. - [03:28](https://www.youtube.com/watch?v=PZHBW6yld5U&t=208s) Performing YouTube competitive scans and drafting titles and thumbnails during the prep phase helps validate viewer demand before committing to recording an interview. - [03:28](https://www.youtube.com/watch?v=PZHBW6yld5U&t=208s) Standard medium-tier reasoning models in ChatGPT provide sufficient depth for multi-step content research and drafting without excessive token costs. - [06:42](https://www.youtube.com/watch?v=PZHBW6yld5U&t=402s) Using the Riverside MCP inside ChatGPT enables AI to scan transcripts, draft intro reels, flag technical issues, and populate task tickets in Linear for human video editors. - [09:31](https://www.youtube.com/watch?v=PZHBW6yld5U&t=571s) High-quality newsletter drafts require explicit human steering to inject specific takeaways and prevent generic summaries. - [09:31](https://www.youtube.com/watch?v=PZHBW6yld5U&t=571s) YouTube thumbnail and title generation requires cross-referencing top channel benchmarks and dedicated human curation, even when assisted by competitive AI scans. - [11:59](https://www.youtube.com/watch?v=PZHBW6yld5U&t=719s) Browser automation tools can handle tedious administrative tasks in publishing platforms, such as copying show notes, enabling monetization, and setting video playlists. - [13:09](https://www.youtube.com/watch?v=PZHBW6yld5U&t=789s) Snapshot diffing captures the exact stylistic gap between raw AI drafts and creator edits, creating an iterative prompt-refinement mechanism. - [14:19](https://www.youtube.com/watch?v=PZHBW6yld5U&t=859s) Social distribution skills perform best when seeded with past winning posts as few-shot examples and paired with specific asset prompts like targeted video frame grabs. - [16:43](https://www.youtube.com/watch?v=PZHBW6yld5U&t=1003s) Modular skill architecture divides workflows into specialized sub-skills coordinated by an orchestration skill that handles asset routing and sequencing. - [16:43](https://www.youtube.com/watch?v=PZHBW6yld5U&t=1003s) The goal of content automation is eliminating mechanical execution while keeping human taste in the loop to prevent the generation of low-quality AI slop. - [18:59](https://www.youtube.com/watch?v=PZHBW6yld5U&t=1139s) Long-term creator leverage comes from building grounded, repeatable workflows around core tools rather than constantly chasing novel, single-purpose AI demos.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=PZHBW6yld5U&t=0s) How I Save 5 Hours a Week with AI

Peter Yang introduces a three-step system he uses to automate his entire content workflow with AI, saving him at least five hours every week. He publishes a new podcast episode every Sunday, a process that previously required him to manually research the guest, prepare an interview guide, record the conversation, write instructions for his video editor, and package everything into a newsletter post, a YouTube video, social posts, and clips. Even with basic AI assistance, this used to take multiple days. Now, he uses a system of interconnected AI skills that automate 90% of the process while still allowing him to apply his own taste and judgment along the way.

Before diving into the walkthrough, Peter gives a quick shout-out to the video's sponsor, Riverside, the all-in-one AI studio he has used for years to record, edit, and produce his podcast at studio-level quality and schedule clips across platforms. Viewers can try it at `creators.riverside.com/peteryang` using the code `PETERYANG` for one month free.

```mermaid flowchart LR accTitle: Three-Phase Content Automation System A["Phase 1: Map every manual step of a workflow"] --> B["Phase 2: Work with AI to create skills for each step"] B --> C["Phase 3: Connect skills together to automate the workflow"] ```

[01:34](https://www.youtube.com/watch?v=PZHBW6yld5U&t=94s) Step 1: Map Each Step of Your Manual Workflow

The first phase of the system is to meticulously map out every single manual step in your existing workflow. Peter shares his own painful 15-step manual podcast production list, noting that even trying to tackle all 15 steps with ChatGPT in a single conversation takes forever. Looking at the list, his workflow naturally breaks down into three distinct phases: - **Preparation:** Researching the guest and putting together an interview guide. - **Editing:** Sending edit instructions to his video editors after recording the episode. - **Post-production:** Getting the transcript for the podcast, uploading the episode to YouTube, drafting title and thumbnail combinations, writing a newsletter post, making clips, writing social posts, and more.

Peter emphasizes that you should not miss any manual steps when building your list: "The more instructions and context you can give AI, the more likely it is to help you automate everything." To demonstrate, he copies his entire 15-step list into ChatGPT (using incognito mode so his existing custom skills don't influence the output) and asks: *"What work from here can you take off my plate?"* ChatGPT responds with a breakdown showing it can own researching the guest, drafting the interview guide, finding the intro reel, cleaning the transcript, making the titles and thumbnails, writing the newsletter post, and making clips—essentially owning most of the workflow with just a little bit of guidance from Peter.

[03:28](https://www.youtube.com/watch?v=PZHBW6yld5U&t=208s) Step 2: Build AI Skills for Those Steps (Podcast Prep)

Step two is creating dedicated AI skills based on the mapped steps. Peter shows the architecture of the AI skills he built to streamline his podcast production, which features one main skill for each of the three phases, plus granular packaging sub-skills called one by one during post-production:

```mermaid flowchart TD accTitle: Podcast Production AI Skills Architecture subgraph Phases["Core Phase Skills"] Prep["podcast-prep (Interview guide)"] Edit["podcast-edit (Video editor instructions)"] Prod["podcast-production (Master packaging skill)"] end subgraph SubSkills["Individual Packaging Skills"] S1["Thumbnail & Title"] S2["Show Notes"] S3["Newsletter Post"] S4["Social Posts"] S5["Video Clips"] end Prep --> Edit Edit --> Prod Prod --> S1 Prod --> S2 Prod --> S3 Prod --> S4 Prod --> S5 ```

To show how these work in practice, Peter walks through a real example: producing his podcast episode *"How to Save Money Now with ChatGPT Finances"* featuring Ethan, the product lead at OpenAI. Inside ChatGPT—where he notes he usually sticks to the standard medium reasoning model ("you don't have to go to Pro or something crazy to save on tokens, so medium is plenty good enough")—he starts with a simple prompt: *"Hey, I'm interviewing Ethan from OpenAI about personal finance. Please do research and help me prep an interview guide."*

This automatically triggers his `podcast-prep` skill, which outputs several key components: - **Recommended end goal & research links:** It determines the interview should focus on practical personal finance use cases and compiles examples and links to research on ChatGPT finances. - **YouTube niche scan & early packaging:** It scans YouTube for well-performing videos in the personal finance and ChatGPT niche and suggests thumbnail and title combinations right off the bat. Peter explains why doing this upfront is critical: "If there isn't any interesting videos about how to use ChatGPT to manage your finances, then maybe this whole interview is actually not worth doing in the first place." - **Initial interview guide:** It drafts an introduction, questions, and notes with example answers based on its research, grouped into three main categories: what ChatGPT can actually do with your money, OpenAI's finance benchmarks (which Peter notes "is not that interesting"), and building a personal CFO that people can trust based on Ethan's background.

Rather than accepting the first draft, Peter gives ChatGPT feedback: look up ChatGPT Finance docs or Ethan's tweet history, focus on use cases, and structure the interview around explaining what ChatGPT Finance is, how to connect accounts, and specific use cases. The `podcast-prep` skill revises the guide into a clean, three-section document (questions, notes, and use cases to demonstrate), which Peter then has ChatGPT write into a Google Doc to send over to Ethan and his team for review.

[06:42](https://www.youtube.com/watch?v=PZHBW6yld5U&t=402s) Recording with Riverside & Running the `podcast-edit` Skill

With prep complete, Peter records the 50-minute interview with Ethan using Riverside. Because Riverside records everything locally, the video remains high resolution without any risk of disconnection artifacts. While Riverside's UI offers built-in AI tools to automatically cut dead air and repeated phrases, Peter prefers to install the **Riverside MCP** directly inside ChatGPT so he can instruct the AI to prepare the podcast for his video editors.

Back in ChatGPT, Peter tells it to run the `podcast-edit` skill for the Riverside Ethan recording. The skill executes a multi-step editorial workflow: 1. **Creates a Linear ticket:** It automatically generates a ticket in Linear containing instructions for his video editing team. 2. **Builds the editor package & intro reel:** It pulls in the thumbnail and title from earlier and drafts a 30-to-40-second intro reel highlighting the most engaging moments of the episode. Using the written transcript, the skill proposes two variations—Option A and Option B. Peter reviews them, gives feedback to combine the two into the exact intro he wants, and the AI updates the Linear ticket. 3. **Extracts clip candidates:** Using the auto-generated written transcript, the skill pulls out compelling standalone clip candidates—for example, a segment on how to audit 90 days of spend with ChatGPT—complete with both the social copy and the full transcript. Peter gives feedback on exact start and stop points, iterates with the AI, and tells it to update the Linear ticket with the finalized clips.

Viewing the finished Linear ticket, everything is organized for his editing team: the thumbnail and title, the intro reel with exact quote locations, a "moments to cut" section where the AI automatically scanned the entire episode to flag logistical issues or technical difficulties with exact timestamps, and the copy and timestamps for all social clips. Listening to the whole podcast again to manually pick out clips and cuts would have taken hours; instead, Peter lets AI scan the transcript while he applies his editorial taste via quick feedback loops.

[09:31](https://www.youtube.com/watch?v=PZHBW6yld5U&t=571s) Drafting the Newsletter Post and Designing Packaging

Moving into podcast production—which constitutes the lion's share of the total workload and is where AI saves the most time—the process kicks off with drafting a Substack newsletter post for every single episode. A typical newsletter entry includes the embedded episode video, a headline, exact timestamps, sponsor acknowledgments, key takeaways, and the most insightful highlights from the conversation. To automate the initial draft, the primary `podcast production` skill calls the dedicated `podcast post` skill. In an episode featuring guest Ethan, the initial AI-generated draft was mediocre: the prompts Ethan discussed weren’t represented accurately, and the framing missed the mark. Rather than settling for weak output, Peter instructed the model to explicitly incorporate the six exact prompts Ethan shared for using ChatGPT to save money, resulting in a much more practical and actionable newsletter draft.

```mermaid flowchart TD accTitle: Podcast Production Workflow accDescr: Diagram showing how podcast production coordinates newsletter creation, packaging, and show notes. subgraph ProductionCore["Podcast Production Orchestration"] ProdSkill["Podcast Production Skill"] end

subgraph ContentAssets["Asset Generation"] ProdSkill -->|"1. Draft"| PostSkill["Podcast Post Skill (Substack Draft)"] ProdSkill -->|"2. Package"| PackSkill["Thumbnail & Title Skill (Packaging & A/B Tests)"] ProdSkill -->|"3. Metadata"| NotesSkill["Podcast Show Notes Skill (YouTube / Spotify)"] end

subgraph ExternalTools["Integrations & Automation"] PackSkill -.->|"Direct edits"| Figma["Figma MCP"] NotesSkill -.->|"Copy & paste / settings"| BrowserUse["Browser Use Tool"] end ```

Once the newsletter copy is in solid shape, the next phase focuses on YouTube packaging via the `thumbnail title` skill. Packaging is crucial for attracting viewers, and this skill functions by scanning the transcript for core themes and conducting automated research on high-performing videos across Peter's channel as well as competing creators. The skill proposes five different thumbnail and title combinations. Even with AI assistance, refining the final packaging typically takes Peter 30 to 40 minutes because he is "super anal about it" and insists on high standards. After several iterations, he locks in the final titles and thumbnail concepts for A/B testing. For creators with established visual templates, this step can be accelerated by connecting tools like a Figma MCP (Model Context Protocol) to let AI manipulate the thumbnail templates directly.

- Newsletter drafts require explicit feedback loops (such as inserting the exact 6 prompts Ethan shared) to prevent generic summaries. - Packaging generation combines transcript analysis with competitive research across top-performing niche channels. - Thumbnail creation can be tied directly into design tools via Figma MCP once visual templates are established.

---

[11:59](https://www.youtube.com/watch?v=PZHBW6yld5U&t=719s) Automating Show Notes and Browser-Based Publishing

The production pipeline next invokes the `podcast show notes` skill to create the episode descriptions for YouTube and Spotify. The skill digests the video timestamps provided by Peter's human video editors, integrates sponsor mentions, adds links back to the Substack post, and automatically sources links to the guest's online profiles. Because the initial pass rarely gets everything right on the first shot, Peter refines the timestamps with feedback—specifically prompting the AI to label the exact use cases covered at each interval.

Once the show notes are polished, Peter uses a browser automation tool (`browser use`) to push the metadata straight into YouTube Studio. The agent copies the text, pastes it into the description box, toggles video monetization on, assigns the video to the appropriate playlist, and configures the rest of the publication settings. Automating these tedious administrative clicks removes a significant chunk of manual friction from the post-production cycle.

---

[13:09](https://www.youtube.com/watch?v=PZHBW6yld5U&t=789s) Iterative Skill Refinement via Snapshot Diffing

When the newsletter post still required manual polish that the AI couldn't nail on the first few passes, Peter implemented a prompt-engineering technique designed to continuously improve AI skills over time. Instead of simply making edits in isolation, he asked the model to take a digital snapshot of its current draft before he touched it. He then applied his manual edits and instructed the AI to compare the before-and-after versions to highlight precisely what changed.

```mermaid flowchart LR accTitle: Skill Improvement Feedback Loop accDescr: Three-step feedback loop comparing AI drafts to human edits to update instructions. Snap["1. Snapshot AI Draft"] --> Edit["2. Human Manual Edits"] Edit --> Diff["3. Generate Diff & Numbered Changes"] Diff --> Update["4. Update Skill Prompt Instructions"] ```

The underlying objective is to refine the `podcast post` skill so it can "one-shot" the task in future episodes without repeating past errors. Peter asked the AI to summarize all the changes between the snapshot and his edited version into a numbered list so he could accept or decline each individual stylistic or structural modification. After reviewing the list and providing notes, the updated guidelines were baked directly back into the skill's system prompt. This structured diffing loop creates a compounding feedback mechanism that steadily eliminates repetitive manual revisions across production cycles.

- Snapshot diffing enables the model to see the concrete stylistic gap between its draft and the creator's taste. - Numbered change proposals give the creator granular control to accept or reject specific guideline updates. - Incorporating human edits back into prompt instructions trains the skill toward true "one-shot" output over time.

---

[14:19](https://www.youtube.com/watch?v=PZHBW6yld5U&t=859s) Crafting Social Posts and Automated Clip Publishing

With the primary long-form assets locked down, the workflow turns to social distribution and short-form video clips using the `podcast social` skill. Peter prompts the agent to write two teaser posts and two long-form podcast posts. To ensure voice consistency, the skill is pre-loaded with high-performing examples of past posts and explicit style rules. Even so, human taste remains necessary to select quotes, adjust hooks, and ensure quality. For the visual asset, Peter asks the AI to find a video frame where both he and Ethan are smiling:

> "This one, Ethan is not smiling, so I want to find one where we're both smiling and pretending we're having fun and then attached it to the teaser post."

The agent locates the smiling screen capture, pairs it with the approved teaser copy, and schedules the post for the upcoming Saturday.

```mermaid flowchart TD accTitle: Social Distribution Pipeline accDescr: Flowchart illustrating how social teasers and video clips are generated, edited, and scheduled. subgraph Teasers["Social Teaser Generation"] SocialSkill["Podcast Social Skill"] --> Hooks["Draft 2 Teasers & 2 Posts"] Hooks --> ScreenCap["Grab Smiling Screen Cap"] ScreenCap --> SchedTeaser["Schedule Teaser"] end

subgraph Clips["Short-Form Video Clips"] ClipPrompt["Select Clip Ideas (e.g., Option 3 Banter)"] --> RiverMCP["Riverside MCP (Trim pauses & words)"] RiverMCP --> Typefully["Typefully Scheduler (Clip + Social Copy)"] end ```

For short-form clips, the AI analyzes the transcript and pitches distinct clip angles, such as executing tax strategies with ChatGPT, finding hidden credit card rewards, or conversational banter. Peter selects the banter option, prompting the system to generate the clip using the Riverside MCP. The Riverside integration cuts out awkward pauses and filler words to deliver a tight clip. After Peter reviews the resulting video and its accompanying caption copy, the agent sends both directly into Typefully—Peter's social scheduling platform—completing the multi-channel distribution pipeline.

- The `podcast social` skill uses past winning posts as few-shot style examples to craft teasers. - Riverside MCP automates rough video editing by trimming dead air, filler phrases, and awkward silences. - Typefully integration handles final scheduling of video clips paired directly with optimized captions.

---

[16:43](https://www.youtube.com/watch?v=PZHBW6yld5U&t=1003s) Architecture of AI Skills and the Automation Framework

Peter summarizes his philosophy: AI should remove repetitive toil while keeping the human in the loop to maintain high standards of taste and craft, preventing the system from producing low-quality "slop." For anyone wanting to automate their own work, the framework consists of three repeatable steps:

1. **Map every manual step**: Document every action in your existing workflow (for Peter's podcast, this totaled 15 distinct steps across three phases). 2. **Build dedicated AI skills**: Feed the mapped steps into an LLM (like ChatGPT or Claude) alongside connected plugins and MCPs to write explicit instruction files. 3. **Connect and orchestrate**: Chain the modular skills together into an end-to-end post-production routing system.

Examining the actual underlying skill files reveals that they look intricate and intimidating, but Peter didn't hand-code most of the text. Instead, he fed conversational instructions and workflow outlines to the AI to generate the markdown documentation defining each skill's trigger conditions, inputs, step-by-step procedures, and routing logic. In the `podcast production` skill, an orchestration layer governs exactly when downstream skills (like packaging, show notes, and clip creation) are called. Peter concludes with a candid reflection on delegation: while AI chaining saves hours, the next logical level of leverage is hiring a human operator to steer the AI agent workflows rather than running all the prompts himself.

- A comprehensive 15-step manual workflow map serves as the foundation for automated skill design. - AI skills contain structured sections: skill name, when to invoke it, precise execution steps, and routing rules. - Orchestration skills act as traffic controllers, routing transcripts and intermediate assets between specialized sub-skills.

[18:59](https://www.youtube.com/watch?v=PZHBW6yld5U&t=1139s) Practical AI Workflows vs. Tool Hype and Closing Resources

Peter wraps up the breakdown by offering a grounded perspective on content creation and automation. He admits that even with this "fancy system setup," it still takes a substantial amount of time and effort to produce his podcast. The intention behind demonstrating his setup is not to pretend that automation solves everything overnight or to chase every shiny new product on the market. Instead of "just doing a bunch of demos or trying a bunch of new tools," Peter wanted to showcase how he realistically uses ChatGPT day-to-day to cut down friction, reclaim hours, and get substantive work out the door.

To help viewers implement these practices without starting from scratch, Peter points to his resource site, noting that anyone looking to adopt these skills and workflows can visit behindthecraft.com. There, he has published a collection of copy-and-paste prompts directly modeled after the system shown in the video. He signs off by thanking viewers and encouraging them to like and subscribe for future "practical no-bullshit tutorials and interviews."

- **Realistic Expectations:** Even an automated pipeline requires dedicated manual effort and editorial oversight to produce high-quality podcast episodes. - **Utility Over Novelty:** Focus on integrated workflows that actually save time on recurring tasks rather than constantly testing disconnected AI app demos. - **Resource Link:** Prompt templates and complete workflow guides are available at `behindthecraft.com` for direct copy-and-paste use.

## Notable quotes > [01:34](https://www.youtube.com/watch?v=PZHBW6yld5U&t=94s) The more instructions and context you can give AI, the more likely it is to help you automate everything. — Peter Yang

> [01:34](https://www.youtube.com/watch?v=PZHBW6yld5U&t=94s) What work from here can you take off my plate? — Peter Yang

> [03:28](https://www.youtube.com/watch?v=PZHBW6yld5U&t=208s) you don't have to go to Pro or something crazy to save on tokens, so medium is plenty good enough — Peter Yang

> [03:28](https://www.youtube.com/watch?v=PZHBW6yld5U&t=208s) Hey, I'm interviewing Ethan from OpenAI about personal finance. Please do research and help me prep an interview guide. — Peter Yang

> [03:28](https://www.youtube.com/watch?v=PZHBW6yld5U&t=208s) If there isn't any interesting videos about how to use ChatGPT to manage your finances, then maybe this whole interview is actually not worth doing in the first place. — Peter Yang

> [09:31](https://www.youtube.com/watch?v=PZHBW6yld5U&t=571s) super anal about it — Peter Yang

> [14:19](https://www.youtube.com/watch?v=PZHBW6yld5U&t=859s) This one, Ethan is not smiling, so I want to find one where we're both smiling and pretending we're having fun and then attached it to the teaser post. — Peter Yang

> [18:59](https://www.youtube.com/watch?v=PZHBW6yld5U&t=1139s) just doing a bunch of demos or trying a bunch of new tools — Peter Yang

> [18:59](https://www.youtube.com/watch?v=PZHBW6yld5U&t=1139s) practical no-bullshit tutorials and interviews — Peter Yang

## Mentioned Peter Yang, Ethan, OpenAI, ChatGPT, Claude, Riverside, Riverside MCP, Linear, Substack, YouTube, YouTube Studio, Spotify, Figma, Figma MCP, Browser Use, Typefully, Google Docs, behindthecraft.com

## Video description 🎙️ I use Riverside to record, edit, and produce my podcast. Try it free for 30 days with code PETERYANG: https://creators.riverside.com/PeterYang

I use ChatGPT and Riverside to save at least five hours a week on my content workflow. In this tutorial, I show how I built 8 AI skills to research guests, prepare interview guides, brief my editors, write newsletter posts, package YouTube videos, and schedule social posts and clips.

What I cover: (00:00) How I save 5 hours a week with AI (01:34) Step 1: Map each step of your manual workflow (03:28) Step 2: Build AI skills for those steps (09:33) Step 3: Connect your skills together (13:09) Improve your AI skills with every edit (14:25) Turn one episode into 6 different content pieces (17:08) How to apply my system to automate your own work

Where to find me: X: https://x.com/petergyang LinkedIn: https://www.linkedin.com/in/petergyang/

Subscribe to this channel - more tutorials coming soon!

[Source](https://www.youtube.com/watch?v=PZHBW6yld5U)

中文简报

Peter Yang 串联 8 项 ChatGPT Skill，削减 90% 播客制作工作量让 AI 将自身草稿与人工修改进行比对，并把变更重新沉淀进提示词指令中，正是训练模块化技能消除重复修改、同时避免产生劣质内容的关键。 Yang 详细拆解了他的端到端播客自动化系统：首先梳理出包含 15 个步骤的人工工作流，针对前期准备、剪辑和后期制作创建专用的 AI 技能，并将它们编排为一个集成流水线。他以采访 OpenAI 的 Ethan 的实际节目为例，展示了 ChatGPT 技能如何与 Riverside MCP、Linear、浏览器自动化以及 Typefully 等工具对接，从而生成采访指南、剪辑师交接清单、缩略图、节目单（show notes）并安排社交平台切片的发布排期。至关重要的是，他强调了人在回路（human-in-the-loop）的把关机制，并介绍了一种“快照比对”（snapshot-diffing）技术来对比 AI 初稿与人工修改，持续训练技能以实现一次成型的准确度，同时杜绝劣质 AI 内容的生成。 Peter Yang 开发了一套由自定义 ChatGPT 技能和 MCP 集成构成的互联系统，实现了每周 15 步播客工作流程中约 90% 的自动化，每期节目至少节省 5 个小时。 (0:00) 通过模型上下文协议（MCP）连接外部服务（例如用于音视频自动粗剪的 Riverside 和用于剪辑工单派发的 Linear），对话式 AI 得以直接操作媒体文件并分派项目任务。 (6:42) 为了系统性地消除重复修改稿件的情况，Yang 采用了一种“快照比对”反馈闭环，让 ChatGPT 将初始输出与人工修改进行比对并生成更新后的提示词规则，从而随着时间推移实现一次成型的准确度。 (13:09) 内容多平台复用由一个主编排技能统一协调，该技能可触发专门的子技能以生成 Substack 文章、YouTube 包装物料、节目单，以及分发至 Typefully 等排期工具中的社交媒体切片。 (9:31)

Reading summary

Peter Yang presents a three-phase automation framework that utilizes chained ChatGPT skills and external tool integrations to handle roughly 90% of his podcast production workflow, saving an estimated five hours per week while preserving human editorial oversight. The Three-Step Workflow Architecture (1:34) The automation framework starts by documenting every task within an existing process before delegating it to AI. Yang mapped his podcast routine into 15 discrete manual steps, which divided naturally into three core stages: interview preparation, video editing instructions, and post-production packaging. Rather than attempting to run a single unmanageable prompt, he designed modular AI skills tailored to specific tasks, such as generating show notes or drafting newsletter entries. To construct these skills, Yang provided ChatGPT with his workflow map and conversational guidelines, allowing the model to generate its own structured markdown instruction files. These files define trigger conditions, execution sequences, and hand-off rules. A central orchestration skill, `podcast-production`, acts as a routing controller, systematically invoking specialized sub-skills as assets move through the production line. Guest Preparation and Video Editor Hand-Off (3:28) The workflow begins with the `podcast-prep` skill, which researches prospective interviewees, identifies high-performing YouTube videos in the subject niche, and suggests early title and thumbnail angles. Yang notes that testing packaging concepts upfront helps determine whether a topic is worth recording at all. After refining the topic focus with human feedback, the model formats the interview questions into a review document for the guest. Following recording on Riverside, Yang triggers the `podcast-edit` skill via a Riverside MCP integration. The AI reviews the transcript to generate a structured Linear ticket for human video editors. This ticket includes proposed 30-to-40-second intro reels, candidate timestamps for short-form clips, and a flagged list of technical glitches or dead air to cut. This automated audit removes the need for Yang to re-listen to raw footage while still allowing him to adjust start and stop points. Automating Asset Packaging and Multichannel Publishing (9:31) The heaviest phase of the pipeline generates public-facing promotional assets. Substack newsletter drafts are assembled by the `podcast post` skill, which pulls guest-specific takeaways and exact interview prompts from the transcript. For video packaging, the `thumbnail title` skill scans channel benchmarks and competitor data to suggest A/B test candidates, an area where Yang spends 30 to 40 minutes refining visual ideas and titles before finalizing them. To handle distribution, the `podcast show notes` skill creates platform-compliant descriptions with sponsor links and editor timestamps. A browser automation tool (`browser use`) transfers this metadata directly into YouTube Studio, handling tagging, monetization toggles, and playlist assignments. Concurrently, short-form video clips selected by the AI are cleaned of filler words via Riverside MCP and forwarded to Typefully alongside scheduled social teaser copy. Snapshot Diffing for Continuous Skill Improvement (13:09) To minimize repetitive manual editing over successive episodes, Yang uses a prompt-engineering feedback loop termed snapshot diffing. When an initial output requires refinement, he instructs ChatGPT to take a snapshot of its draft before applying his manual edits directly to the text. The model is then prompted to run a diff between the original snapshot and the edited draft, summarizing the stylistic and structural revisions into a numbered list. After Yang approves or rejects specific adjustments, the confirmed guidelines are integrated directly back into the skill's core system instructions. This iterative refinement trains the system toward reliable 'one-shot' generation over time. Yang emphasizes that automation does not replace human taste or eliminate the effort required to make quality media, warning against chasing novel AI tools without practical workflows. He notes that while chaining AI agents significantly cuts friction, the next step in scaling leverage is handing prompt execution to a human operator.

Source quotes

[1:34] Peter Yang: The more instructions and context you can give AI, the more likely it is to help you automate everything. [1:34] Peter Yang: What work from here can you take off my plate? [3:28] Peter Yang: you don't have to go to Pro or something crazy to save on tokens, so medium is plenty good enough [3:28] Peter Yang: Hey, I'm interviewing Ethan from OpenAI about personal finance. Please do research and help me prep an interview guide. [3:28] Peter Yang: If there isn't any interesting videos about how to use ChatGPT to manage your finances, then maybe this whole interview is actually not worth doing in the first place. [9:31] Peter Yang: super anal about it [14:19] Peter Yang: This one, Ethan is not smiling, so I want to find one where we're both smiling and pretending we're having fun and then attached it to the teaser post. [18:59] Peter Yang: just doing a bunch of demos or trying a bunch of new tools [18:59] Peter Yang: practical no-bullshit tutorials and interviews

Original transcript

[0:00] Today, I'm going to show you how I use [0:01] AI to automate an entire content [0:04] workflow to save me at least five hours [0:06] a week. By the end of this video, you'll [0:08] learn how to copy my exact three-step [0:11] system to automate your own manual [0:13] workflows. I'm confident that you'll be [0:15] able to use AI to take back at least a [0:17] few hours of your time after watching [0:19] this. All right, so I publish a new [0:22] podcast episode every Sunday. And for [0:25] each episode, I used to have to manually [0:27] research the guest, prepare an interview [0:29] guide, record the conversation, write [0:31] instructions for my video editor, and [0:33] package the whole thing into a [0:35] newsletter post, a YouTube video, social [0:37] posts, and clips. And this would take me [0:40] multiple days to do, even with AI's [0:42] help. But now, I use a system of AI [0:45] skills that work together to automate [0:48] 90% of this process while still letting [0:51] me apply my taste and judgment along the [0:53] way. This system has three phases. [0:56] First, map every manual step of a [0:59] workflow. Then, work with AI to create [1:02] skills for each step. And finally, [1:04] connect your skills together to automate [1:07] the entire workflow. Now, real quick, [1:09] this tutorial is brought to you by [1:10] Riverside, the all-in-one AI studio that [1:12] I use to record, edit, and produce my [1:15] podcast. I've been using it for years, [1:17] and I love how it records at studio [1:19] level quality and lets me edit and [1:20] schedule clips to all platforms. You can [1:23] try yourself by going to [1:24] creators.ai.com/peteryang [1:28] and use code peteryang to get one [1:30] month's free. All right. Now, let me [1:32] show you exactly how I work with AI. So, [1:34] as I mentioned, the first step is to map [1:37] every step of a manual workflow. Here is [1:41] everything that I used to do manually to [1:43] produce my podcast. You can see that [1:45] there are 15 steps which you can imagine [1:47] is pretty painful to do and even working [1:49] with chat GPT and AI to tackle all these [1:52] steps in a single conversation just [1:54] takes forever. So just from making this [1:56] list you can see that there's three [1:58] distinct phases to my workflow. First I [2:01] have to prepare the podcast by [2:03] researching the guest and then putting [2:06] together an interview guide. Then after [2:08] recording the episode, I have to send [2:10] edit instructions to my video editors. [2:14] And last but not least, I have to do [2:15] postp production, which is getting the [2:17] transcript for the podcast, getting the [2:20] episode uploaded to YouTube, drafting [2:22] title and thumbnail combinations, [2:24] writing a newsletter post, making clips, [2:26] writing social posts, and more. So, you [2:29] should also put together a very [2:30] meticulous list like this. Don't miss [2:32] any manual steps. The more instructions [2:34] and context you can give AI, the more [2:36] likely it is to help you automate [2:38] everything. So in fact, why don't we [2:39] actually just copy this whole list into [2:41] chatbt and ask it what work from here [2:46] can you take off my plate and let's see [2:49] what it comes up with. Okay, here I'm [2:51] using a cognto mode because I've already [2:52] built skills for a lot of stuff and I [2:54] don't want those to impact the output. [2:57] All right, so here is chatb suggestions. [3:00] It's saying that it can own researching [3:02] the guest, drafting the interview guide. [3:05] It can own finding the intro real. And [3:08] then down here, it can say it can own [3:10] clean the transcript, making the title [3:12] and thumbnails, newsletter post, making [3:14] the clips, and so on. Says it can own [3:16] most of it with a little bit of guidance [3:18] from me, which is exactly the flow that [3:20] we want. Okay, so that's step one, which [3:23] is to map every step of your workflow [3:25] and then brainstor AI which parts it can [3:27] automate. Now, let's move on to step [3:29] two, which is creating AI skills based [3:32] on these steps. So, here are all the AI [3:35] skills that I created to streamline my [3:38] entire podcast production workflow. You [3:40] can see that there's one skill for each [3:42] phase. There's podcast prep for making [3:44] the podcast interview guide. There's [3:46] podcast edit for sending instructions to [3:48] my video editor. And then there's [3:50] podcast production, which is the most [3:51] complicated skill, which is about [3:53] packaging. And what this skill does is [3:55] it walks me through these other skills [3:58] to make the thumbnail title, show notes, [4:00] the newsletter post, the social post, [4:02] and even video clips one by one by [4:04] calling these individual skills. So, let [4:07] me actually show you how everything here [4:08] works starting with podcast prep. And [4:11] we're going to use a real example [4:13] throughout, which is how I produced this [4:15] podcast about how to save money now with [4:17] ChachiP finances with Ethan, the product [4:20] lead. All right. So, here I am in chat [4:23] GPT. And by the way, for all these [4:25] workflows, I usually stick to GPT. So, [4:28] medium. You don't have to go to Astra or [4:30] something crazy to save on tokens. So, [4:32] medium is plenty good enough. But I [4:35] start this whole process by saying, [4:36] "Hey, I'm interviewing Ethan from OpenAI [4:38] about personal finance. Please do [4:41] research and help me prep an interview [4:43] guide." And it's automatically going to [4:45] call the podcast prep skill. And this is [4:47] what it's going to come out with. So, I [4:49] did a bunch of research and I found out [4:51] that the recommended end goal for my [4:52] interview with Ethan is to focus on [4:55] practical personal finance use cases. [4:57] Here's some examples. Here's a bunch of [5:00] links to the research that I did about [5:01] chatb finances. And then it also does a [5:05] YouTube scan for wellperforming videos [5:08] in this niche. So, personal financing [5:09] chatbt and so on. And it even suggests [5:12] packaging options right off the bat for [5:14] thumbnail and title combinations that I [5:17] should think about. And the reason it [5:18] does all of this is because, for [5:20] example, if there isn't any interesting [5:21] videos about how to use chat to manage [5:23] your finances, then maybe this whole [5:25] interview is actually not worth doing in [5:27] the first place. So, it's worth doing [5:28] this research up front to figure out if [5:31] this interview is even worth doing. Now, [5:34] it then puts together the actual [5:36] interview guide for Ethan. Here's the [5:38] introduction that I to do. Here is the [5:41] questions and some notes from example [5:43] answers based on his research. So the [5:46] main categories that it has is what chat [5:48] GBD can actually do with your money. [5:50] Something about open eyes finance [5:52] benchmarks which is not that interesting [5:55] and then also building a personal CFO [5:57] that people can trust based on Ethan's [5:58] background. So what I did next was I [6:01] gave it some feedback. I said look up [6:03] Chat GB Finance docs or Ethan's tweet [6:05] history and focus on use cases and [6:08] here's the structure of the interview [6:09] that I actually want. I want to explain [6:11] what chatb Finance is. I want to talk [6:13] about how to connect your accounts. I [6:14] want to talk about use cases and so on [6:16] so forth. And then it revised interview [6:18] guide based on my feedback. All right. [6:21] So here's the final interview guide that [6:23] I iterated on with the podcast prep [6:25] scale. And there's three clear sections. [6:27] There's questions. There's some notes. [6:29] There's some use cases I want to [6:30] demonstrate. And I got Chachi to write [6:33] all this stuff in Google Doc and send it [6:35] over to Ethan and the team for review. [6:37] Okay. So that was step one, which is [6:39] podcast prep, researching and preparing [6:41] the podcast. Next, I recorded the [6:44] interview with Ethan. And once again, I [6:46] use Riverside to record the interview. [6:47] Here is the finished recording. It's 15 [6:49] minutes long. And as you can see, it [6:51] records everything locally, so [6:53] everything remains high resolution, and [6:55] you don't have to worry about getting [6:56] disconnected. It also has a bunch of AI [6:58] tools to help you cut dead air, cut [7:00] repeat phases, and so on so forth [7:02] automatically. [7:04] Okay. But what I like to do is instead [7:06] of using the UI here, I like to use the [7:09] Riverside MCP and install that in Chat [7:12] GPT and then give AI instructions to [7:15] prepare the podcast for my video [7:18] editors. [7:19] So, let me show you how that works. [7:22] Okay, so we're back in ChattP and I told [7:23] it to run the podcast edit scale for the [7:26] Riverside Ethan recording. And the first [7:29] thing it did is it made a linear ticket [7:31] to include instructions for my video [7:33] editors. And then it decided to go ahead [7:36] and make the editor package. And the [7:38] editor package includes the thumbnail [7:40] title that we did before. But it also [7:42] includes the intro reel. So every [7:44] episode I make has a short 30 to 40 [7:47] second intro reel that highlights the [7:49] most engaging stuff in the podcast. My [7:51] skill basically puts together two [7:53] different intro reels based on the [7:55] written transcript of the podcast. just [7:57] option A and option B. And then I gave [8:00] it some feedback to combine the two. And [8:02] this is the intro that I want. Then it [8:05] went off and it put it into the linear [8:07] ticket. And next, it also decided to [8:10] make some clip candidates too. So again, [8:13] it's using the automatically generated [8:14] written transcript from the podcast and [8:16] it's pulling out interesting clip [8:18] candidates. So the first one is about [8:21] how you can audit nine days of spend [8:23] with catchupt and it writes both the [8:26] social copy and also the full [8:27] transcript. [snorts] So I gave it some [8:29] feedback again to improve the clips to [8:32] tell it where to start and where to stop [8:34] and then it's doing some more [8:35] iterations. [8:37] And then I said okay now update the [8:39] linear ticket with all this stuff [8:41] including the clips. So let's take a [8:43] look at the linear ticket now. All [8:44] right. So, this is the linear ticket for [8:46] my video editing team. And there's a few [8:49] different sections that we work through. [8:50] There's the thumbnail and title. There's [8:52] the intro reel and where to pull the [8:54] quotes. There is moments to cut. So, [8:56] automatically it went through the entire [8:58] episode and it found stuff to cut [9:00] logistic issues or technical [9:02] difficulties and it gave the timestamps [9:04] for this. And then it also has the copy [9:07] and the stuff for all the clips that I [9:09] want to make. So again, for me to do [9:11] this stuff manually, listen to the whole [9:12] podcast again and pick out interesting [9:14] moments would take hours. But instead, [9:16] I'm using AI to look at the entire [9:18] transcript, find interesting moments. [9:20] I'm still applying my taste along the [9:22] way. I'm just not just letting it spit [9:23] out his output and pasting it here. I'm [9:25] giving a feedback to make it good, but [9:28] it just saves me a lot of time. All [9:29] right, so that is the podcast edit part [9:31] of the process. And now let's talk about [9:34] podcast production, which is the [9:36] majority of the work and also where AI [9:38] saves me the most time. For every [9:40] episode that I make, I also create this [9:42] newsletter post for my Substack. And it [9:44] has the video, it has the title, it has [9:47] the timestamps from the video, the [9:50] sponsor, and also takeaways and stuff [9:53] that's super insightful from the podcast [9:56] that you can see here. Okay. So to [9:59] create the newsletter post, I basically [10:01] use podcast production to call the [10:03] podcast post skill as the first draft. [10:06] So here's the example. I say, "Hey, [10:07] let's work on a newsletter post." For [10:09] this one, I wanted to be super [10:10] practical, focus on the actual prompts [10:12] that Ethan shared. And here it drafted [10:15] the first version of the newsletter [10:16] post. Now, the thing is, the first [10:18] version was not that good. There was a [10:19] lot of issues. The prompts weren't quite [10:21] right. It didn't quite use what I want. [10:23] And I told it, "Hey, these are the exact [10:25] six prompts that Ethan shared with us. [10:28] Can you please incorporate these prompts [10:30] into the newsletter post? Okay. And [10:33] based on the feedback, you can see here [10:34] in the newsletter post that I went off [10:36] and pasted the exact prompts that Ethan [10:38] shared for how to use chat GPT to save [10:41] money. So with the newsletter post in [10:44] relatively good shape, I next told it to [10:47] run the thumbnail title scale to figure [10:49] out what the packaging is. And what this [10:51] skill does is it looks through the [10:53] transcript again. It finds out what are [10:55] the main topics discussed during the [10:57] podcast and then it does a bunch of [10:59] research on outperforming videos both [11:02] from my channel and other similar [11:05] channels. And based on this research, [11:08] it suggests different combinations for [11:10] the thumbnail and title. Now, if you are [11:13] a YouTuber, you know how important the [11:14] packaging is. Find the right thumbnail [11:16] and title. It's just really important [11:17] for people to come and actually watch [11:18] the video. So, here's five different [11:21] packages that is suggested for me. And I [11:24] thought about it and I said, "Hey, why [11:25] don't we do these combinations [11:28] instead?" And then we went back and [11:30] forth and here is the thumbnails I came [11:33] up with. And by the way, once you have a [11:35] repeatable thumbnail template like I [11:37] have for my podcast, you can also just [11:39] tell AI to use Figma MCP or whatever [11:42] tool you use to edit a thumbnail [11:44] directly. Usually, even with AI's help, [11:46] it takes me at least like 30 to 40 [11:48] minutes to come up with the right [11:49] thumbnail title just because I'm like [11:50] super anal about it and I want to make [11:51] it good. Okay, so here's the [11:53] combinations that we come up with and [11:55] that we are going to AB test for this [11:57] episode. All right, so going back to [11:59] this map here, uh I also have a skill [12:01] called podcast show notes that helps me [12:03] draft YouTube and Spotify descriptions. [12:07] So basically I told you to hey draft the [12:09] show notes for this for me and here is [12:11] the proposed show notes that came up [12:12] with the right timestamps based on the [12:15] uploaded YouTube video that my editors [12:17] did for me. It also has links to [12:19] sponsors, links to my newsletter post, [12:21] and where to find Ethan that it found [12:24] itself online. A best practice is if you [12:27] find yourself repeating the same [12:28] template or the same process each time, [12:30] just build a skill to streamline as much [12:32] of it as possible. And just like with [12:34] the newsletter post, this stuff isn't [12:35] perfect, unfortunately, in one shot. So, [12:37] I give it feedback to make it better. [12:39] For example, name the actual use cases [12:41] in the time stamps and iterations there. [12:45] And then I told it to add the show notes [12:48] to the YouTube description itself. So it [12:50] uses browser use to copy what we agreed [12:53] on and paste it here. And it can also [12:56] click all these buttons to turn on [12:57] monetization to make sure that it's part [12:59] of the right playlist and so on and so [13:01] forth, which again is just like a lot of [13:04] manual work that you can streamline with [13:07] AI. All right. Now, going back to the [13:09] newsletter post, I wasn't fully happy [13:11] with what AI came up with. So I decided [13:14] to make a bunch of manual edits to the [13:16] post directly. And here's something that [13:18] I think all of you can learn from. So [13:21] before I made the manual edits to the [13:22] post, I asked it to take a snapshot of [13:25] the current post that had AI's edits and [13:28] then I'm going to make manual edits next [13:30] and then I want you to compare the two. [13:32] So it took a snapshot. Then I made my [13:34] edits and then it found that here's what [13:37] changed. And and the reason I told it to [13:38] explain all this is because I wanted to [13:40] update the scale so that it can oneshot [13:42] and avoid these mistakes the next time. [13:45] So you can see here the next prompt I [13:46] gave it is summarize all the before and [13:48] after changes to the podcast post scale [13:51] to be able to oneshot this next time and [13:54] list these changes in a number list so [13:57] that I can accept or decline each [13:58] change. So it went off and listed [14:00] everything in a number list here and I [14:03] actually read through and reviewed [14:04] everything and gave a bunch of feedback [14:06] and then it given me a new list and I [14:08] said hey make all the updates to the [14:10] scale. So this is like a pretty common [14:12] process that I use to improve my skills [14:14] over time so that it doesn't repeat the [14:16] same mistakes over and over. All right, [14:19] so going back to this map, we have the [14:21] thumbnails, we have the titles, we have [14:23] the show notes, and we have the [14:24] newsletter post. So, next step is to [14:26] make a social post and also prepare and [14:29] publish clips. So, to make the social [14:32] post, I have my podcast social skill. I [14:34] tell it to give me two teasers and two [14:36] podcast posts. So, here's a bunch of [14:38] teasers that I wrote. And I like how you [14:40] can click around to figure out which [14:42] teaser I like. And also, here is the [14:44] podcast post. And by the way, I came up [14:46] with this stuff because I give each [14:48] skill like really good examples of past [14:51] teaser and podcast post that I wrote and [14:54] tell her to follow the same style and [14:55] the same guidelines. But you can see [14:57] again it doesn't make it perfect in one [14:59] shot. So I have to make manual edits. I [15:01] can tell you to pick a better intro [15:03] quote and so on so forth. And then I [15:06] told her, okay, I'm pretty happy with [15:07] this. Why don't you go find a good [15:09] screen cap from the YouTube video where [15:12] we're both smiling. This one, Ethan is [15:14] not smiling, so I want to find one where [15:16] we're both smiling and pretending we're [15:17] having fun and then attached it to the [15:20] teaser post. Okay, so it went off and it [15:22] found this screen cap. Here we go. It [15:25] looks like we're both having a good [15:26] time. Then it attached it to the [15:29] schedule post. So, let's go ahead click [15:30] on this. You can see here that this [15:32] teaser post is now scheduled for next [15:35] Saturday with the copy we agreed on and [15:38] also with the image that AI found. All [15:40] right, so last but not least, AI can [15:42] also help you find interesting clips [15:43] from the podcast and schedule them to [15:46] social media. And to do this, I just say [15:48] video post. Now, let's make a clip. Give [15:51] me your top suggestions. Make chat GPT [15:54] execute your tax plan number one, find [15:56] hidden credit card rewards number two, [15:59] and so on and so forth. And down here, I [16:02] told it, hey, I really like option three [16:05] because of the banter that we have. and [16:07] go ahead and make the clip with [16:09] Riverside. Let's take a look. So, if you [16:12] use the Riverside MCP again and then [16:14] made the clip here for me to review. And [16:16] you notice how I cut out a bunch of [16:18] pauses and other unnecessary words to [16:21] make the clip that I actually want. So, [16:23] I reviewed the clip and I also reviewed [16:25] the social post that goes along with the [16:27] clip and I said, "Looks good. Go ahead [16:29] and put both on Typefully, which is my [16:31] social scheduling app to schedule the [16:34] post." So, let's go ahead and review the [16:36] post in Tyfully. And here we have the [16:38] social post along with the full clip in [16:41] Typhoia as well. All right. So, there [16:43] you go. That's a super quick walkthrough [16:45] of how I built a series of AI skills to [16:48] streamline my entire podcast production [16:51] process. Now, as I'm showing you, I'm [16:54] still really involved in each step to [16:56] review AI's output, to make manual edits [16:58] if necessary, because I don't want it to [17:00] just pump out slop. I want it to be high [17:02] taste and high craft. But these skills [17:04] and chain them together really helped me [17:06] save a ton of time. So just to recap, if [17:09] you want to do this for yourself, [17:10] because I know not everyone has a [17:11] podcast, is just pick one or two manual [17:15] workflows that are taking up a lot of [17:17] your time every week. And then the first [17:19] step is to just list all the manual [17:22] steps in each workflow. So here's 15 [17:24] manual steps from my podcast process. [17:27] Then just ask AI, ask or ask claude, [17:30] hey, here's all the manual steps I take. [17:32] which parts of this can you help me [17:33] streamline? And the more mcps and [17:37] plugins you connect chatpt and cloud to, [17:39] the more likely it is to streamline more [17:41] of these steps. Then once you list the [17:43] manual steps, there's clear patterns [17:45] that emerge. There's three different [17:46] phases to my podcast process. So just [17:49] build the skills with AI for each phase. [17:53] And by the way, I want to actually show [17:54] you what these skills look like. So for [17:56] example, this is the podcast prep skill [17:57] that I use to prepare the interview [18:00] guide. And this looks like pretty [18:01] complicated and pretty intimidating, but [18:03] to be honest with you, I didn't write [18:04] most of this. I gave AI instructions and [18:07] I came up with a workflow. So here's the [18:09] name of the scale. Here's when to use [18:11] it. And here's how to use it. Exactly. [18:13] Similar thing for podcast edit scale, [18:16] there's a list of steps that AI should [18:18] take. And also for podcast production [18:21] scale there's a list of steps and [18:23] there's a whole orchestration and [18:25] routing layer of different other skills [18:27] it should call based on the sequence [18:29] that I'm in doing post-production for [18:31] the podcast. So the important thing is [18:33] to not get intimidated by all this and [18:35] just remember my three steps map every [18:38] manual step of a workflow. Create AI [18:40] skills based on these steps and connect [18:42] these skills together end to end to [18:45] automate as much of the workflow as [18:46] possible. And real talk, I feel like I [18:50] probably need to be better at actually [18:52] hiring a human to help drive AI to do a [18:55] lot of stuff instead of just doing [18:56] everything myself because even with this [18:59] fancy system setup, I still take a lot [19:01] of time to produce my podcast. But yeah, [19:04] I just want to show you guys how I [19:06] actually do work with chat GPT. Instead [19:08] of just doing a bunch of demos or trying [19:10] a bunch of new tools, this is how I [19:12] actually try to save time and get work [19:13] done. All right, so that's it for this [19:16] episode. I hope you enjoyed it. For more [19:19] of my skills and workflows, you can go [19:20] to behindthecraft.com [19:23] where I've included a bunch of copy and [19:24] paste prompts that you can use. And as [19:27] always, thanks for watching and please [19:30] like and subscribe for more practical [19:32] nobullshit tutorials and interviews.