ChatGPT Unseats Claude Across Seven of Ten Everyday AI Tasks

ChatGPT vs Claude vs Grok vs Gemini: The Best AI for 10 Use Cases (August 2026) — Peter Yang

Peter Yang

OpenAI's desktop agent harness, browser automation, and cross-thread voice orchestration make the $20 ChatGPT plan the single best AI subscription, while Claude Opus 5 loses ground to preachy prose and repetitive writing tics.

AI-assisted summary of ChatGPT vs Claude vs Grok vs Gemini: The Best AI for 10 Use Cases (August 2026) — original source. Check the original for full context.

Source published:

Key ideas

  • After previously relying almost exclusively on Claude, Peter Yang crowns ChatGPT the overall winner across 10 everyday AI use cases—and the best single $20/month subscription—citing its rapid shipping pace, generous rate limits, and superior voice and browser harnesses. (25:52)
  • While Claude Opus 5 has developed a preachy tone and repetitive writing tics, Claude remains the top choice for strategic planning and end-to-end design workflows—spanning clarifying UX questions in Claude Design, animated HTML slide decks, and programmatic video editing via Claude Fable. (3:01)
  • A sharp coding divide has emerged between non-engineers and senior developers: ChatGPT wins for 'vibe coding' by autonomously testing in the browser and sustaining long debugging contexts, whereas engineers inspecting diffs often prefer Grok to avoid 500-line over-engineered outputs. (13:55)
  • In the race for personal AI agents, model weights matter less than the surrounding software harness—where the ChatGPT desktop app currently leads via cross-thread voice orchestration and browser automation, though Grokbot's dedicated 24/7 cloud VMs pose a major competitive threat. (23:55)

Detailed notes

## Background Published on YouTube by Peter Yang on August 26, 2026, this video evaluates ChatGPT, Claude, Grok, and Gemini head-to-head across ten everyday productivity use cases. The description explains that while Yang previously relied almost exclusively on Claude, the rapid diversification of foundation models now requires selecting specialized tools tailored to specific tasks.

## TL;DR Peter Yang conducts a ten-part benchmark evaluating ChatGPT, Claude, Grok, and Gemini across domains including UI design, writing, strategic planning, coding, browser use, voice control, and personal agent harnesses. Claude retains a strong edge in architectural design prompts, HTML presentations, and strategic decision-making, though it suffers from stylistic writing clichés and personality degradation. ChatGPT dominates voice thread orchestration, persistent iterative coding, custom-tailored editing, and browser automation, while Grok and Gemini flash situational strengths in zero-to-one visual layouts and entertaining video clips. Weighing ecosystem flexibility and daily execution, Yang names ChatGPT the decisive overall winner and the best single $20-per-month investment for everyday workflows.

## Atmosphere & style The presentation carries the pragmatic, unvarnished tone of an experienced product builder sharing his personal workstation setup rather than reciting sterile benchmark charts. Yang balances systematic software comparisons with warm, relatable anecdotes—from testing recipe prototypes while visiting family in Canada to laughing at surreal AI ramen commercials with his kids or dictating multi-agent tasks while resting after lunch. Moving at a swift, engaging pace, the walkthrough cuts through corporate AI marketing with blunt evaluations of model tics, sycophancy, and interface friction.

## Bulletin - [25:52](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1552s) After previously relying almost exclusively on Claude, Peter Yang crowns ChatGPT the overall winner across 10 everyday AI use cases—and the best single $20/month subscription—citing its rapid shipping pace, generous rate limits, and superior voice and browser harnesses. - [03:01](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=181s) While Claude Opus 5 has developed a preachy tone and repetitive writing tics, Claude remains the top choice for strategic planning and end-to-end design workflows—spanning clarifying UX questions in Claude Design, animated HTML slide decks, and programmatic video editing via Claude Fable. - [13:55](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=835s) A sharp coding divide has emerged between non-engineers and senior developers: ChatGPT wins for 'vibe coding' by autonomously testing in the browser and sustaining long debugging contexts, whereas engineers inspecting diffs often prefer Grok to avoid 500-line over-engineered outputs. - [23:55](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1435s) In the race for personal AI agents, model weights matter less than the surrounding software harness—where the ChatGPT desktop app currently leads via cross-thread voice orchestration and browser automation, though Grokbot's dedicated 24/7 cloud VMs pose a major competitive threat.

## Key takeaways - [00:32](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=32s) In zero-to-one mobile app prototyping, Grok unexpectedly outperforms ChatGPT, Claude Code, and Gemini in immediate visual impact by combining functional UI layouts with clean, integrated AI-generated imagery. - [03:01](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=181s) Claude Design excels at complex product design because it is the only surface that pauses to ask targeted clarifying questions about target audience, design systems, and UX constraints before generating code. - [04:21](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=261s) For automated HTML slide decks and programmatic video editing via Claude Fable and the open-source Hyperframes skill, Claude is the clear winner—leading to the prediction that AI will replace 90% of professional human video editing within three months. - [07:34](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=454s) Claude's conversational personality peaked with Opus 4.6 and became preachy and judgmental in Opus 5, making ChatGPT (configured with custom anti-sycophancy rules) far better for everyday answers and casual dialogue. - [09:43](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=583s) ChatGPT outperforms Claude in long-form writing and editing because it conforms more reliably to custom style guides and evaluation loops while avoiding repetitive rhetorical 'Claudisms' like 'the uncomfortable reality' or 'this is X, not Y.' - [12:06](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=726s) Claude remains the wisest strategic thinking partner for difficult quarterly planning decisions, catching subtle operational vulnerabilities—such as an upcoming staffing departure and a tendency to hoard manual work—that ChatGPT overlooks. - [13:55](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=835s) ChatGPT is the top coding pick for non-engineers because it persistently tests apps via browser tools and handles long multi-turn debugging sessions without context degradation, whereas senior engineers who read raw code often favor Grok for concise, minimal diffs. - [16:01](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=961s) Browser and computer automation is the most underrated AI category for navigating legacy web portals without APIs—such as government filings or cross-platform social publishing with platform-specific tagging—with ChatGPT's desktop app leading the field. - [19:14](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1154s) ChatGPT's voice mode acts as a hands-free orchestration engine rather than just a chat interface, allowing users to spin up and route instructions across multiple named background threads while away from their desk. - [20:20](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1220s) In complex infographic image generation, ChatGPT beats Grok and Gemini because its conversational memory automatically applies stored brand palettes and visual guidelines while respecting constraints to keep copy light. - [22:31](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1351s) For consumer AI video generation, Gemini easily defeats Grok Imagine by producing entertaining, coherent clips with voiceovers, whereas Grok severely distorts facial features in unsettling ways. - [23:55](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1435s) Winning the personal agent race depends on the software harness rather than raw model weights; ChatGPT's desktop app wins today on browser, voice, and mobile integration, while Grokbot stands out by giving every agent its own dedicated 24/7 cloud virtual machine. - [25:52](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1552s) For users choosing a single $20-per-month subscription, ChatGPT delivers the best overall value across the generative AI landscape, while Claude, Gemini, and Grok serve best as specialized complements for design/strategy, video/Workspace, and cloud agents respectively.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=0s) Overview: Testing ChatGPT, Claude, Grok, and Gemini Across 10 Use Cases

A few months back, Peter Yang's workflow relied almost exclusively on Claude for virtually every task. However, the AI landscape has fractured rapidly, producing a wide array of specialized tools and models where the optimal choice strictly depends on the exact task at hand. To resolve this, Peter sets up a head-to-head showdown evaluating ChatGPT, Claude, Grok, and Gemini across ten core everyday use cases—including writing, UI/UX and visual design, slide generation, and video editing—demonstrating real-time tests for each category to illustrate why particular models pull ahead.

[00:32](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=32s) Use Case 1 (Design): Mobile App Prototyping

Peter starts the competition with design because it serves as the most visually dynamic and challenging testbed. In his definition, design encompasses everything from websites and mobile applications to presentation slide decks and animated launch videos. To ground the test in a real-world scenario, Peter mentions that he is currently visiting his parents in Canada and wants to cook for them. He issues an identical design and UX prototyping prompt across ChatGPT, Claude, Grok, and Gemini:

> *"Design and prototype a mobile app to discover and cook simple recipes. There's two screens, discover recipe screen and a view recipe screen, and focus on the UX only."*

The resulting implementations showcase radically different strengths and weaknesses across the contenders:

- **ChatGPT**: Produces a functional but bare-bones recipe discovery prototype. Tapping an item reveals ingredients and instructions, but Peter notes surprise that ChatGPT did not tap into its native image-generation capabilities, leaving empty image placeholders and an underwhelming presentation. - **Claude Code**: Delivers a far more functional prototype featuring recipe filters, serving size adjustments, and step-by-step cooking directions under the headline *"What are we cooking tonight?"* While mechanically robust, Claude lacks native image generation, leaving visuals absent. It also suffers from a repetitive, recognizable aesthetic: specific header styling above titles, telltale typography, and an overall recognizable "Claude-generated" template feel. - **Grok**: Serves as the biggest surprise of the test. It outputs a clean, visually polished layout populated with matching AI-generated imagery (such as a distinct photo for a "tomato butter pasta" recipe), fully functional buttons, and structured cooking methodologies. Grok wins the initial zero-to-one design round on immediate visual impact. - **Gemini**: Generates an app using image generation alongside Nano Banana to fill out assets, but suffers from messy alignment bugs, odd border artifacts, and scattered text placement. While the individual recipe view looks cleaner than its home screen, Gemini trails behind Grok.

```mermaid flowchart TD accTitle: Mobile App Prototyping Comparison accDescr: Comparison of prototype outputs across Claude, Grok, ChatGPT, and Gemini.

subgraph Comp["Model Prototyping Results"] GrokNode["Grok: Best 0-to-1 visual prototype with integrated AI imagery"] ClaudeNode["Claude Code: Deep functionality and UX filters, but lacks images and has signature styling"] GPTNode["ChatGPT: Basic functional prototype with no image generation"] GeminiNode["Gemini: Generated images, but plagued by layout and border glitches"] end ```

[03:01](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=181s) Deep-Dive on Claude Design: The Power of Clarifying Questions

While Grok claims the zero-to-one visual trophy, Claude offers a specialized surface called **Claude Design**. Submitting the same recipe prompt into Claude Design triggers a crucial onboarding step that none of the competing models attempt: the model pauses to ask targeted architectural and product design questions before generating code. It queries Peter on the target audience (which he sets to *"busy people with 20 minutes"*), the specific design system requirements, and what *"discover"* should practically mean inside the interface.

Once answered, the interface streams in tailored components that strictly respect those constraints. Although the UI continues to exhibit Claude's signature visual palette and misses out on AI-generated food photography (forcing it to rely on blank placeholders or potential future integrations like Unsplash), the preliminary clarification workflow enables Claude to follow complex product directions far more accurately than its rivals.

[04:21](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=261s) HTML Slide Decks and Claude Fable Video Generation

Moving beyond standard app wireframes, Peter explores automated slide deck generation. He completely eschews building manual slides in PowerPoint or Google Slides, writing all presentations via HTML rendered with AI. Testing a deck on personal agent workflows using his custom slide skill and brand guidelines:

- **ChatGPT** generates a deck titled *"Rockbot is the future of personal agents"*, incorporating a six-step usage list and basic structural diagrams. However, it injects an unprompted, nonsensical circle and outputs a fairly rigid, basic diagram. - **Claude** executes the exact same slide design prompt with significantly higher production value. Claude's HTML slide output features animated, twinkling stars on diagrams, cleanly removes the weird glitchy shapes seen in ChatGPT's version, and animates individual UI icons. Peter highlights Claude as his default engine for building decks.

The visual showdown concludes with **Claude Fable**, which Peter utilizes for end-to-end video production and editing workflows via the open-source **Hyperframes** skill from GitHub:

- **Full Product Launch Videos**: Using text prompts alone, Claude Fable 5 produces a completely animated launch video for Peter's *Human Review* skill, featuring smooth camera scrolls, screen transitions, and real dynamic pacing that resembles a genuine SaaS launch asset rather than a glorified slideshow. - **Talking-Head Video Overlays**: With a custom *Video Intro* skill, Claude Fable watches raw talking-head footage, automatically transcribing captions, identifying contextual moments, sourcing icons autonomously, and executing visual zooms and on-screen overlays.

Peter anticipates that automated AI video editing is roughly three months away from replacing 90% of what a human professional editor does. Claude remains the definitive overall winner for comprehensive design spanning websites, mobile apps, presentation slides, and programmatic video—provided the prompter explicitly instructs it to avoid its standard beige background and red accent look.

[07:34](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=454s) Use Case 2: Everyday Answers and Personality

Shifting from visual creation to routine dialogue and daily queries, Peter highlights that a model's intrinsic voice and tone become the deciding factors. A few months prior, chatting with Claude felt like conversing with an intimate, trusted friend. However, that rapport has degraded. Peter reflects that **Opus 4.6** represented the high-water mark for Claude's personality. By comparison, **Opus 5** exhibits an annoying, preachy tone, becoming overly judgmental and repeatedly falling back on canned corporate idioms like *"Here's the honest truth."*

Consequently, Peter has migrated away from Claude for casual chatter in favor of **ChatGPT**, which recently improved by stripping out its single most grating habit: appending sycophantic suggestions to the bottom of every turn, such as *"If you want, I can also do X. I can also answer question Y."* Grok and Gemini, while competent, are disqualified from this category for being far too stiff, verbose, and mechanical.

To combat the inherent sycophancy, excessive hedging, and boilerplate phrasing present across all foundation models out of the box, Peter shares three mandatory custom rules he inputs under ChatGPT's Personality settings:

- *"Be candid and thoughtful. Tell Peter what he needs to hear instead of what you think he likes to hear."* - *"Write in the active voice. Minimize hedging, padding, and grand claims."* - *"Favor specific and concise responses. Make every word earn its place. And never use AI slop words like delve, foster, leverage, and so on and so forth."*

[09:43](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=583s) Personality Wrap-Up and Writing & Editing

Wrapping up the previous evaluation on personality, ChatGPT takes the win because of its less judgmental tone and far more generous rate limits on the standard $20 per month plan, avoiding sycophantic praise and annoying AI filler words. Moving into writing and editing, the landscape represents another dramatic reversal. Claude's writing quality has visibly devolved over recent months, picking up frustrating mannerisms dubbed "Claudisms"—formulaic rhetorical tics like "the uncomfortable reality" or "this is X, not Y" that become grating in long-form prose. ChatGPT, by comparison, conforms far more reliably to a specific personal writing voice and instructions.

To force AI models to write naturally, the key is supplying explicit style instructions and evaluation loops. Using a custom "newsletter skill" inside ChatGPT, draft blog posts are passed through a structured workflow containing a voice-and-tone style guide drawn from real writing samples. Rather than generating content entirely from scratch, the tool edits an existing draft, details every change made, and grades its own output across voice, hook, substance, and payoff. It also integrates a standalone "no AI slop" filter—an open-source system designed to strip out cliché phrases and bloated sentence structures, particularly common in Claude.

```mermaid flowchart LR accTitle: Newsletter Skill Editing Workflow accDescr: A visual diagram tracing a draft through the newsletter skill, scoring criteria, and slop filtering. A["Raw First Draft"] --> B["Newsletter Skill"] subgraph Review ["Rules & Evaluation"] B --> C["Voice & Tone Guide"] B --> D["No AI Slop Filter"] end Review --> E["Edited Draft + Change Log"] E --> F["Self-Score (Voice, Hook, Substance, Payoff)"] ```

- **Writing Tics:** Claude heavily leans into predictable phrasing patterns ("the uncomfortable reality," "this is X not Y") that require active pruning. - **Slop Prevention:** The standalone "No AI Slop" prompt and skill repository has gained over 5,000 GitHub stars for cleaning up repetitive generative tropes. - **Behind The Craft:** Peter offers his complete library of handcrafted skills, a setup course, and a private community for automating personal AI systems at `behindthecraft.com`.

---

[12:06](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=726s) Planning and Difficult Strategic Decisions

When faced with high-stakes strategic choices—such as deciding which paid offer to launch, which active project to sunset, or what objectives to prioritize across a three-month horizon—Claude remains the preferred thinking partner. Despite its quirks in prose style, Claude currently stands out as the most thoughtful, wise, and macro-aware reasoning model on the market for parsing nuance.

To test this, both ChatGPT and Claude were given a raw planning document and asked to recommend key focuses for the next quarter. While ChatGPT offered sensible, surface-level advice (building out the personal AI course and focusing on user activation and renewals), Claude went a level deeper. Beyond validating the core membership launch, Claude identified a subtle operational vulnerability that ChatGPT completely overlooked: closing an impending video operations gap. An operations assistant handling video clips and post-production was scheduled to depart on August 19th. Claude recognized that without hiring an operational replacement or virtual assistant cheaply and quickly, Peter would fall back into his known habit of trying to do all the execution alone or via AI instead of delegating.

- **Claude's Edge in Strategy:** Claude excels at catching contextual subtext and operational landmines buried inside planning notes. - **The Delegation Blind Spot:** The model flagged Peter's specific tendency to hoard manual workflow execution rather than backfilling administrative personnel. - **Current Standing:** Claude holds the planning crown, at least until next-generation reasoning releases (such as OpenAI's upcoming models) hit the market.

---

[13:55](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=835s) Coding: Practical Grit vs. Clean Architecture

The pick for coding comes with an admitted caveat: building as a non-engineer who ships small apps and open-source tools without inspecting raw lines of code. For this "vibe coding" style, the process begins by structuring requirements with Claude, but the actual execution is handed off entirely to ChatGPT (specifically GPT-4o). ChatGPT approaches software tasks with persistent grit—it "has a dog in it." It will spin up browser tools, test the application autonomously, diagnose interface breakages, and iterate through bug fixes until the feature works end-to-end without needing manual intervention.

ChatGPT also handles sprawling, multi-turn coding chats far better than Claude. Where Claude frequently runs into memory bloat and requires manual prompt context compaction to avoid degradation, ChatGPT sustains deep, continuous debugging loops cleanly. However, experienced software engineers often take a drastically different view. A former L8 engineer friend, Kun, points out that both GPT-4o and Claude Opus heavily over-engineer solutions, churning out bloated, 500-line diffs when only a targeted fix is needed. For engineers who actually read the code, Kun strongly prefers Grok for delivering concise, minimal code modifications.

- **Coding Persona:** ChatGPT aggressively tests and debugs via built-in tooling, refusing to quit until an implementation functions end-to-end. - **Context Handling:** Claude degrades faster in lengthy coding chats, whereas ChatGPT maintains coherence across long conversational histories without frequent manual context pruning. - **The Engineer's Counter-Pick:** Professional developers who inspect diffs often favor Grok over GPT-4o or Opus to avoid unnecessary architecture and sprawling code generation.

---

[16:01](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=961s) Browser and Computer Use: The Most Underrated AI Frontier

Browser and desktop automation represents perhaps the most overlooked category in everyday productivity. Vast amounts of corporate and bureaucratic knowledge work remain trapped inside legacy web portals that either lack APIs or charge exorbitant rates for programmatic access. Delegating browser navigation allows models to scrape, fill forms, and interact with the web just like a human operator.

A striking example comes from an OpenAI employee who used ChatGPT's browser automation to prepare an entire multi-year immigration filing package in minutes, extracting seven years of tax records, bank statements, and required paperwork. Peter uses a similar pattern for cross-platform social distribution. Using a custom social posting skill, a single command drafting a quote tweet about a Marie Kondo Grok bot was automatically converted and published across X, LinkedIn, Threads, and Substack. The model is intelligent enough to adapt to platform rules—tagging the specific `@GrokBot` account on X while omitting the handle on Threads and LinkedIn where that account does not exist.

```mermaid flowchart TD accTitle: Browser-Driven Social Automation Flow accDescr: Diagram illustrating a single social skill publishing across multiple platforms with automated platform-specific tagging. A["Raw Prompt + Media"] --> B["Social Post Skill"] B --> C{"Platform Dispatch via Browser"} C -->|"Includes @GrokBot Tag"| D["X (Twitter)"] C -->|"Strips Non-Existent Handle"| E["LinkedIn & Threads"] C -->|"Formats Long-Form Text"| F["Substack"] ```

- **High-Friction Portals:** Browser automation is ideal for tedious government documents, healthcare portals, complex file uploads, and even letting AI click through mandatory corporate training videos. - **Model Comparison:** ChatGPT’s desktop application leads browser navigation. Claude possesses computer-use capabilities but feels less integrated; Grok is viable on dedicated personal machines; Gemini remains an unfinished work in progress. - **Desktop Workflow:** The sheer convenience of native browser execution is the primary reason the ChatGPT desktop client remains an essential daily tool.

---

[18:52](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1132s) Voice Chat and Agent Orchestration

ChatGPT maintains a decisive lead in voice capabilities, driven by low latency and fluid conversational dynamics. Rather than using voice merely for simple queries, the ChatGPT desktop app functions as an audio-driven command center to orchestrate background threads, manage ongoing tasks, and coordinate external agents hands-free while working across multiple desktop environments.

[19:14](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1154s) Voice Chat: Cross-Thread Orchestration on ChatGPT

Voice mode in ChatGPT shines brightest not merely as a conversational tool, but as a hands-free orchestration engine that can command other background workflows. Peter demonstrates this by issuing an ambitious command entirely by voice: asking the assistant to spin up a newsletter thread inside his "Personal OS" to edit a recent post, check in on open action items within his "chief of staff" thread, and inspect whatever remaining tasks exist in a "video intro skill" thread. Because voice requires referencing existing workflows by name, maintaining a clean, structured naming convention across threads becomes essential. The system parses the instructions, launches a dedicated editing thread, and routes instructions across the designated channels.

Peter explains that this workflow is the ultimate hack for getting things done when you want to be "ultra lazy." Instead of sitting at a desk typing out instructions, he frequently goes for an afternoon walk or lies down in bed after a heavy lunch, talking into ChatGPT voice while the AI coordinates his various threads behind the scenes. Because none of the competing models match this seamless level of background thread orchestration and mobile responsiveness, ChatGPT takes the voice category by a wide margin.

```mermaid flowchart TD accTitle: Voice Thread Orchestration Flow accDescr: Diagram showing how a single ChatGPT voice command routes tasks to multiple specialized threads. V["Voice Command in Personal OS"] --> T1["Edit Newsletter Post Thread"] V --> T2["Chief of Staff Thread"] V --> T3["Video Intro Skill Thread"] T1 --> O1["Drafts & Edits Recent Post"] T2 --> O2["Follows Up on Open Items"] T3 --> O3["Audits Video Intro Tasks"] ```

[20:20](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1220s) Image Generation: Infographics Across ChatGPT, Grok, and Gemini

To evaluate image generation, Peter puts the models through a complex layout challenge: creating an infographic from the full text of his newsletter post titled *"How we use computers is about to change forever."* Claude is immediately sidelined because it lacks built-in image generation capabilities, leaving ChatGPT, Grok, and Gemini to compete. The prompt instructs each tool to transform the post into an infographic while keeping the copy light and clean.

The results vary wildly across the three contenders:

- **Grok:** Delivers a disappointing, low-effort graphic dominated by generic clip-art illustrations. Peter considers it easily the weakest and most basic output of the test. - **Gemini (Imagen / Nano Banana):** Produces a noticeably more detailed infographic where the diagrams and structure make sense. However, it fails to obey the negative constraint to keep text sparse, resulting in cluttered layouts. It also injects strange artifacts—such as a literal "nano banana"—partly because Peter mentioned the model tag in his prompt. - **ChatGPT (DALL-E):** Takes the top spot due to instruction adherence and conversational memory. Because ChatGPT retains context from past conversations about Peter's visual brand guidelines, it automatically adheres to his preferred brand palette, circular motifs, and illustration aesthetic. While the image elements are slightly oversized, the overall execution follows instructions far better than the alternatives.

[22:31](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1351s) AI Video Generation: Grok Imagine vs. Gemini for Crazy Japanese Commercials

Video generation currently remains a two-horse race among the major chat ecosystems, as neither Claude nor ChatGPT offers integrated end-to-end video generation within their standard consumer chat interfaces. To test Gemini and Grok Imagine, Peter uses one of his favorite family pastimes: uploading a photo of himself and asking the AI to produce a wildly eccentric Japanese television commercial promoting delicious ramen, complete with high-energy visuals and a Japanese voiceover.

Gemini delivers an impressive and entertaining clip that hits the comedic brief perfectly, provoking laughs from Peter and his kids—even if he jokes that he will never publish it publicly just to "save face." In contrast, Grok Imagine produces an unsettling result. It wildly distorts Peter's facial features and shows him erratically jumping around in a way that feels "pretty damn scary." While Peter notes that Grok likely maintains looser content moderation filters than Google—and specialized Chinese generators like SeaArt operate with virtually no guardrails—Gemini easily wins this round for consumer-friendly, entertaining video generation.

[23:55](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1435s) Personal Agents: Harnesses, Tooling, and the Race That Matters Most

When evaluating personal AI agents—systems designed to complete multi-step tasks across external tools and everyday apps—Peter argues that model weights matter far less than the software "harness" wrapping around them. The race comes down to interface flexibility, execution environments, and ecosystem hooks:

- **ChatGPT Desktop App:** Peter's daily driver. It leads the industry in browser automation, voice fluidity, and cross-platform mobile access. Although the interface can feel cluttered when juggling multiple threads across modes like Codex, Chat, and Work, it remains the most capable and flexible harness available. - **Claude Code:** Anthropic's promising developer-centric harness, though it still trails OpenAI in general browser operation, mobile integration, and voice controls. - **Grokbot:** Offers a notably cleaner interface than ChatGPT and introduces an impressive architecture where every agent receives its own dedicated 24/7 cloud virtual machine. Its main limitation is structural rigidity: users cannot easily spin up multiple chat threads with the same agent. - **Gemini Spark:** Google's personal agent tool integrates tightly with Google Workspace and has begun rolling out third-party integrations, but lacks the deep library of advanced plugins available on ChatGPT.

Peter awards the personal agent category to ChatGPT, though he notes that if Grokbot introduces multi-thread flexibility for its cloud-hosted agents, it could become a serious threat. In his view, winning the personal agent harness is the single most critical battleground in consumer AI—outweighing even raw coding benchmarks—because the platform that masters cross-app agentic workflows will ultimately capture "the lion's share of consumer attention."

[25:52](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1552s) Overall Winner and the $20-a-Month Recommendation

Reviewing his complete comparison grid across all ten use cases, Peter crowns ChatGPT as the definitive overall champion. With OpenAI shipping updates to its desktop harness, voice tools, and coding assistants at an aggressive pace, competitors face an uphill battle to dethrone it as an all-in-one platform.

Each rival retains distinct strengths for specialized needs: Claude continues to excel in design conceptualization and strategic planning; Gemini holds the edge in rapid video generation and Google Workspace integrations; and Grok is emerging as a serious player through Grokbot's dedicated agent architecture. Nevertheless, for anyone looking to maximize their productivity without juggling multiple subscriptions, Peter concludes that paying $20 a month for ChatGPT delivers the absolute best value across the entire generative AI landscape. He closes by directing viewers to behindthecraft.com for his comprehensive system on setting up personal AI workflows.

## Notable quotes > [00:32](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=32s) Design and prototype a mobile app to discover and cook simple recipes. There's two screens, discover recipe screen and a view recipe screen, and focus on the UX only. — Peter Yang

> [07:34](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=454s) Here's the honest truth. — Peter Yang

> [07:34](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=454s) If you want, I can also do X. I can also answer question Y. — Peter Yang

> [07:34](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=454s) Be candid and thoughtful. Tell Peter what he needs to hear instead of what you think he likes to hear. — Peter Yang

> [07:34](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=454s) Write in the active voice. Minimize hedging, padding, and grand claims. — Peter Yang

> [07:34](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=454s) Favor specific and concise responses. Make every word earn its place. And never use AI slop words like delve, foster, leverage, and so on and so forth. — Peter Yang

> [09:43](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=583s) the uncomfortable reality — Peter Yang

> [09:43](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=583s) this is X, not Y — Peter Yang

> [13:55](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=835s) has a dog in it — Peter Yang

> [19:14](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1154s) ultra lazy — Peter Yang

> [20:20](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1220s) How we use computers is about to change forever. — Peter Yang

> [22:31](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1351s) pretty damn scary — Peter Yang

> [23:55](https://www.youtube.com/watch?v=nAhQs8Fd_9g&t=1435s) the lion's share of consumer attention. — Peter Yang

## Mentioned Peter Yang, OpenAI, ChatGPT, GPT-4o, DALL-E, Anthropic, Claude, Claude Code, Claude Design, Claude Fable, Opus 4.6, Opus 5, Grok, xAI, Grokbot, Grok Imagine, Google, Gemini, Gemini Spark, Imagen, Nano Banana, Hyperframes, Behind The Craft, Kun, PowerPoint, Google Slides, Unsplash, X, LinkedIn, Threads, Substack, SeaArt, Google Workspace

## Video description Copy my best AI workflows to save time and automate busywork: https://www.behindthecraft.com 🤝 Subscribe to my practical AI newsletter read by 150K+ professionals: https://creatoreconomy.so

A few months ago, I used Claude for almost everything. Today, there are so many AI models to choose from that the best really depends on your use case.

In my new tutorial, I compare ChatGPT, Claude, Grok, and Gemini head-to-head for writing, design, coding, personal agents, and 6 other use cases.

What I cover: (00:00) ChatGPT, Claude, Grok, and Gemini across 10 use cases (00:32) Still the best model for design (apps, slides, video) (07:34) This model's personality just isn't the same (09:58) The best model for writing and editing (12:06) The model I trust for difficult decisions (13:55) My controversial pick for coding (16:01) The most underrated AI category (18:52) Why ChatGPT voice is still far ahead (20:20) Which model follows image prompts best? (22:31) Grok vs. Gemini for AI video (23:55) The AI race that matters more than coding (25:52) My overall winner (and the best model to start with)

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=nAhQs8Fd_9g)

中文简报

OpenAI赢得2026年AI对决：ChatGPT在十大用例中力压竞争对手 ChatGPT在十大用例中摘得综合桂冠，其在语音聊天、浏览器自动化和多线程编排上的主导优势超过了Claude在设计与规划方面的长处。 Peter Yang在真实任务中对四款前沿AI平台进行了系统测试，评估了它们在界面设计、战略规划、日常写作、编程、浏览器自动化和多媒体生成方面的表现。Claude在复杂的幻灯片与应用设计以及精细的商业战略规划方面保持领先；Gemini在创意新奇的视频生成中拔得头筹；Grok则在从零到一的UI原型制作中展现出潜力。总体而言，凭借无可比拟的语音编排、适合非工程师的高适应性自主编程以及稳健的桌面浏览器自动化，ChatGPT成为无可争议的赢家和最佳日常主力工具。 ChatGPT在一项包含10个类别的基准测试中成为综合赢家，在写作、非技术性编程、语音编排、浏览器自动化以及个人代理工作流方面斩获最高排名。 (25:52) Claude在宏观战略规划和界面设计方面保持领先，通过发现其他模型忽略的运营漏洞展现出卓越的上下文综合能力。 (12:06) 个人AI代理的前沿竞争主要由软件适配框架（software harnesses）而非底层基础模型驱动，这使得桌面执行环境比代码生成基准更为关键。 (23:55) 自主浏览器与计算机操作成为最被低估的AI能力，它通过将受困于缺乏标准API的老旧门户系统的工作流实现自动化，带来了巨大的生产力提升。 (16:01)

Reading summary

Evaluating ChatGPT, Claude, Grok, and Gemini across ten everyday workflows, Peter Yang argues that the AI landscape has fractured by task—shifting his daily reliance from Claude to ChatGPT as the best all-around tool. Tone, Writing Quality, and Strategic Planning (7:34) Yang reports a sharp reversal in conversational quality and prose between Claude and ChatGPT. While he views Claude Opus 4.6 as the former high-water mark for personality, he criticizes Opus 5 for becoming preachy and judgmental, as well as relying on formulaic "Claudisms" like "the uncomfortable reality" or "this is X, not Y." ChatGPT now wins both everyday dialogue and editing because it removed sycophantic follow-up suggestions, offers more generous rate limits on its $20 monthly plan, and follows custom style rules. Yang forces natural writing in ChatGPT using a custom "newsletter skill" that scores drafts on voice, hook, substance, and payoff while applying an open-source "No AI Slop" filter (which has over 5,000 GitHub stars) to strip words like "delve" and "leverage." Despite its prose quirks, Claude remains Yang's preferred partner for three-month strategic planning. When analyzing a quarterly planning document, ChatGPT offered sensible surface-level advice on course building and renewals, whereas Claude caught a critical operational blind spot: an operations assistant handling video post-production was departing on August 19th, and failing to hire a replacement would trigger Yang's habit of hoarding execution himself. Visual Design, Prototyping, and Multimedia Generation (0:32) In visual creation, performance splits between raw asset generation and structured product design. When prototyping a two-screen recipe app, Grok unexpectedly won zero-to-one visual impact by pairing a polished layout with AI-generated food photography, whereas ChatGPT left empty image placeholders and Gemini suffered from alignment and border glitches. However, Claude won the overall design category through specialized surfaces. Unlike rivals that immediately generate code, Claude Design pauses to ask clarifying questions about target audiences and design systems, though users must instruct it to avoid its recognizable default beige-and-red aesthetic. Claude also outperformed ChatGPT in HTML slide generation by adding clean animations, and Yang uses Claude Fable (with the open-source Hyperframes skill) to generate animated SaaS launch videos and automated talking-head overlays—forecasting that AI is roughly three months away from replacing 90% of a human video editor's work. For standalone media generation, Claude is sidelined by its lack of native image and video tools. ChatGPT won an infographic test on Yang's newsletter by adhering to negative constraints for light text and remembering his brand palette from past chats, whereas Grok produced basic clip-art and Gemini cluttered the layout with text and a literal "nano banana" artifact. Conversely, in a consumer video test generating an eccentric Japanese ramen commercial from a selfie, Gemini delivered an entertaining clip while Grok Imagine unsettlingly distorted Yang's face. Coding Approaches: Vibe Coding vs. Engineering (13:55) The optimal coding model depends heavily on the user's technical background. Operating as a non-engineer who ships small apps without inspecting raw code, Yang outlines requirements in Claude but hands execution to ChatGPT (specifically GPT-4o). He notes that ChatGPT autonomously spins up browser tools to test interfaces, fix bugs end-to-end, and sustain long multi-turn debugging chats without the memory bloat and manual context compaction required in Claude. However, Yang contrasts his workflow with the perspective of Kun, a former L8 software engineer who actually reads code diffs. According to Kun, both GPT-4o and Claude Opus heavily over-engineer solutions—often generating bloated 500-line diffs for targeted fixes—making Grok the superior choice for professional developers seeking concise, minimal modifications. Browser Automation, Voice, and Agent Harnesses (16:01) Yang contends that the software "harness" wrapping a model matters far more for personal agents than raw model weights, positioning cross-app orchestration as the most critical battleground for consumer attention. ChatGPT's desktop app leads this category through superior browser automation and voice controls. Browser navigation unlocks work trapped in legacy portals lacking APIs; Yang cites an OpenAI employee who compiled seven years of tax and bank records for an immigration filing in minutes, alongside his own workflow that adapts and publishes social posts across X, LinkedIn, Threads, and Substack (automatically stripping the `@GrokBot` handle where that account does not exist). Furthermore, ChatGPT's low-latency voice mode acts as a hands-free command center, allowing Yang to route tasks across distinctively named background threads—such as his newsletter, chief of staff, and video intro threads—while walking or resting. Competing agent harnesses remain more specialized or constrained. Claude Code trails in general browser use, mobile integration, and voice; Gemini Spark integrates well with Google Workspace but lacks deep third-party plugins; and while Grokbot gives every agent a dedicated 24/7 cloud virtual machine inside a clean interface, its inability to spin up multiple chat threads with the same agent currently holds it back. Although Claude excels in design workflows and strategic planning, Gemini leads in consumer video and Google Workspace ties, and Grok shows promise in minimal code diffs and cloud agent architecture, Yang crowns ChatGPT the overall winner. Because of OpenAI's rapid shipping pace across its desktop harness, voice orchestration, and coding tools, he concludes that a single $20-per-month ChatGPT subscription delivers the highest all-around value.

Source quotes

[0:32] Peter Yang: Design and prototype a mobile app to discover and cook simple recipes. There's two screens, discover recipe screen and a view recipe screen, and focus on the UX only. [7:34] Peter Yang: Here's the honest truth. [7:34] Peter Yang: If you want, I can also do X. I can also answer question Y. [7:34] Peter Yang: Be candid and thoughtful. Tell Peter what he needs to hear instead of what you think he likes to hear. [7:34] Peter Yang: Write in the active voice. Minimize hedging, padding, and grand claims. [7:34] Peter Yang: Favor specific and concise responses. Make every word earn its place. And never use AI slop words like delve, foster, leverage, and so on and so forth. [9:43] Peter Yang: the uncomfortable reality [9:43] Peter Yang: this is X, not Y [13:55] Peter Yang: has a dog in it [19:14] Peter Yang: ultra lazy [20:20] Peter Yang: How we use computers is about to change forever. [22:31] Peter Yang: pretty damn scary [23:55] Peter Yang: the lion's share of consumer attention.

Original transcript

[0:00] Hey everyone. So, just a few months ago [0:03] I was only using Claude to do pretty [0:05] much everything. [0:06] But today, there are so many AI tools [0:08] and models out there [0:10] that the best really depends on your use [0:13] case. [0:14] In this video, I'm going to compare [0:15] ChatGPT, Claude, Grok, and Gemini [0:19] head-to-head and pick the winners across [0:21] 10 common use cases that you see here, [0:24] ranging from writing to design and more. [0:27] I'll show live demos of each use case as [0:29] well so you can see why I like the [0:31] winner. [0:33] Okay, so let's start with design because [0:35] it is the most visually interesting use [0:37] case. [0:38] And by design, I mean designing [0:39] websites, apps, slide decks, and even [0:43] animated videos. [0:45] So, I'm visiting my parents in Canada [0:47] and I want to cook for them. [0:49] So, let's ask ChatGPT, Claude, Grok, and [0:52] Gemini to basically do this. Design and [0:54] prototype a mobile app to discover and [0:57] cook simple recipes. There's two [0:59] screens, discover recipe screen and a [1:01] view recipe screen, and focus on the UX [1:05] only. [1:06] All right, let's take a look at [1:07] ChatGPT's output first. [1:09] So, you can see here that it created a [1:10] pretty basic prototype to discover [1:13] recipes and if I click on one of this [1:15] it'll let me see the ingredients and [1:18] steps for a single recipe. But this is [1:20] pretty basic. I'm surprised that it did [1:21] not use his image generation [1:23] capabilities to actually fill in the [1:24] images. [1:26] Now, let's switch over to Claude code. [1:28] So, I put the same prompt here [1:30] and Claude code made this app. Let's [1:32] open it. [1:33] And here it is. What are we cooking [1:35] tonight? There's filters in place. If I [1:37] click on this, there is servings. It [1:40] definitely has more functionality than [1:42] ChatGPT prototype. [1:44] And there's steps. Now, the problem with [1:46] this is that although it looks visually [1:48] great, Claude doesn't really have any [1:49] image generation capabilities. And also, [1:51] this whole design just looks very kind [1:54] of Claude generated. Claude likes to put [1:56] like the header here above the title. It [1:59] likes to use this kind of fonts. And the [2:01] whole thing just looks like very Claude [2:03] generated. And I'll show you how to fix [2:04] this in a little bit. But now let's [2:07] actually take a look at Grok's [2:08] prototype. And this really surprised me. [2:10] Grok created a visually impressive [2:13] prototype. It's clearly laid out. And it [2:15] also uses image generation features to [2:17] fill it in, which makes it look a lot [2:19] better. And the images actually [2:21] correspond to the text. So let's take a [2:22] look at tomato butter pasta. [2:25] And you can see here it opens here. And [2:27] there's ingredients, and there's [2:29] methodology. And there's buttons that [2:31] also work. So far this is the most [2:33] visually impressive cuz Grok has both [2:35] images and a great layout. But let's [2:37] close by looking at Gemini's prototype. [2:40] So this is Gemini. It also use image [2:43] generation and nano banana to fill in [2:45] the images. [2:46] But you can see that there's a lot of [2:47] issues with Gemini's prototype. There's [2:48] a bunch of border and alignment issues. [2:50] There's text all over the place. Let's [2:52] click on a recipe. [2:53] I think this looks a little bit better [2:55] or organized. [2:56] But I think just from a pure zero to [2:58] one, Grok wins, right? But the thing [3:01] about Claude is that it also has Claude [3:03] design. [3:04] So let me actually copy this prompt. [3:08] And let me open Claude design. [3:10] And let's go ahead and paste the same [3:11] prompt in Claude design and say mobile [3:14] app design and hit submit. Now the thing [3:16] I really love about Claude design is [3:19] that it's going to ask me some [3:20] clarifying questions first. So it's [3:22] going to ask me any design system to [3:24] build on. Who is cooking? Let's say busy [3:26] people with 20 minutes. [3:29] What should discover actually mean? [3:31] Okay, let's pick some options here. [3:33] And let's send the answer. And this is [3:36] actually a really key step that I'm [3:37] surprised none of these other guys do. [3:39] Because it really helps me kind of [3:41] clarify the designs. And then AI can [3:44] actually follow my instructions more [3:46] precisely. So let's wait a little bit [3:48] and see what Claude design comes up [3:50] with. [3:51] All right. So now we can see the designs [3:53] streaming in. You can see that it [3:54] respected my picks around 20 minutes, [3:57] busy people trying to cook as quickly as [3:59] possible. [4:00] And I wish it didn't look as much like [4:02] Claude, [4:03] but it kind of follows my direction a [4:05] little bit more. And the other thing [4:06] that's a little bit of a drawback is [4:09] because Claude didn't invest in image [4:11] models, a lot of the images are still [4:13] missing. Maybe it can pull images from [4:15] like Unsplash or somewhere else, [4:17] but the design just doesn't look as good [4:19] without the images. [4:21] All right. So, that's the design [4:22] comparison for mobile app screens. Now, [4:24] let's try comparing how GPT and Claude [4:28] does in designing slides. So, these days [4:31] I make all my slides using HTML with AI [4:34] instead of manually in PowerPoint or [4:36] Google Slides. [4:38] And basically, I asked GPT to build this [4:40] deck with my slide skill, and here is [4:43] the copy for the slide. [4:45] And then it went off and applied my [4:47] brand guidelines to build the deck. So, [4:50] here is GPT's deck. Rockbot is the [4:53] future of personal agents. It has a [4:55] little diagram here. [4:56] And there is six ways to get the most [4:58] out of Rockbot. [5:00] And this looks pretty good, but there's [5:01] some random circle here that I didn't [5:03] specify. [5:04] And overall, this diagram looks a little [5:06] bit basic. [5:08] Okay. So, I basically got Claude to do [5:10] the same thing, and this is Claude's [5:12] deck. [5:13] So, very similar because I have a style [5:15] template that it should follow, but I [5:16] think the diagram is a little bit more [5:18] animated. There's like stars appearing. [5:20] And the next slide is also a little bit [5:22] more visually appealing. That weird [5:24] circle is gone. There's even some [5:26] animations in some of these icons here. [5:30] So, whenever I want to make a slide [5:31] deck, I always use Claude to make it, [5:34] and here's the rest of the deck. [5:35] And yeah, it can make you really [5:37] beautiful slides if you give it the [5:39] right copy and use skills like my slide [5:41] skill. So, last but not least, I want to [5:43] show you how I use Claude Fable to [5:46] create and edit videos. And it's pretty [5:49] amazing at this, especially if you give [5:51] it [5:52] this free hyper frames skill that's [5:55] available on GitHub. [5:57] So, let me show you a video that it made [5:59] for me for my human review skill. [6:03] And yeah, this entire product launch [6:04] video was made by Claude Fable 5 just [6:07] with text instructions for me, right? [6:09] This is pretty impressive. It's very [6:11] animated. It kind of scrolls. It changes [6:13] things. It doesn't just look like a [6:14] slide deck. It looks like a real product [6:17] launch video. [6:19] And now, let me show you something else. [6:21] So, I also built another skill called [6:23] video intro skill [6:24] where [6:25] it actually helps me add overlays to my [6:28] talking head videos. So, this is a video [6:30] that I made that I got Claude Fable to [6:32] add captions and other animations to. [6:35] Let's play it. [6:36] So, you can see it's adding captions. [6:38] It's adding overlays here. It's adding [6:40] icons that it found by itself. [6:42] And it can also support effects like [6:45] zoom and other things, too. And yeah, [6:48] this isn't quite replace a professional [6:49] video editor, but I was pretty amazed [6:52] that it's able to actually help me [6:53] animate and illustrate videos like this. [6:57] I think AI video editing is maybe 3 [6:59] months away from doing 90% of what a [7:01] professional editor does. [7:03] And I'll make a dedicated video about [7:05] how to use Claude to edit videos in the [7:07] near future. [7:08] So, definitely like and subscribe uh to [7:10] follow along. [7:11] All right, so we just spent a lot of [7:12] time just talking about this first [7:14] design use case because I feel really [7:15] passionate about it. [7:17] So, overall, I think Claude still wins [7:18] on designing websites, apps, slides, and [7:20] videos. [7:22] But, Grok is actually a pretty good [7:23] contender, especially for websites and [7:25] apps. With Claude, you just have to make [7:27] sure to give it instructions to avoid [7:29] the typical Claude kind of beige [7:31] background, red highlight look. So, [7:34] that's for design. Now, let's talk about [7:36] everyday answers. [7:38] And for everyday answers, I I the [7:40] model's personality really matters. [7:43] So, just a few months ago, you know, [7:44] talking to Claude felt like talking to a [7:47] close friend. It had the best [7:48] personality. But, something's been off [7:51] since then. [7:52] And to be honest, I feel like Opus 4.6, [7:54] which they launched way back then, was [7:57] the best Claude model for personality. [7:59] Opus 5, in contrast, [8:01] it tends to be like a little bit too [8:02] judgemental, and it tends to use [8:04] Claude-speak like, "Here's honest [8:06] truth." a little bit too much. [8:08] So, I don't enjoy talking to Claude [8:10] anymore, if I'm honest. [8:12] And instead, I now prefer to just talk [8:15] to ChatGPT. So, ChatGPT has become a lot [8:19] more interesting to talk to because it [8:21] got rid of its most annoying quirk, [8:23] which is at the end of every response, [8:25] it used to say, "If you want, I can also [8:27] do X. I can also answer question Y." [8:29] which is just like super fake and [8:31] annoying. And this was only like 6 [8:33] months ago, but now it's gone. Now, I'm [8:35] not going to demo Grok and Gemini, but [8:38] in my opinion, they tend to be a little [8:39] bit too mechanical and verbose in their [8:41] responses. [8:43] So, overall, I'm going to have to give [8:45] the winner for everyday answers and [8:47] personality to ChatGPT, the new winner [8:50] from Claude. [8:52] But, real quick, honestly, none of these [8:53] models has a super great personality out [8:57] of the box. In my opinion, they're a [8:58] little bit too verbose, sympathetic, [9:01] like to suck up to you, and they're also [9:02] full of AI slop words. So, what I [9:05] recommend that you guys all do is go to [9:07] ChatGPT [9:08] and go to the settings for ChatGPT, [9:11] and then go to personality. And then, in [9:14] here, add some custom instructions. So, [9:16] these are the three pieces of custom [9:18] instructions that I think are really [9:19] important. [9:20] Be candid and thoughtful. Tell Peter [9:22] what he needs to hear instead of what [9:24] you think he likes to hear. Write in the [9:26] active voice. Minimize hedging, padding, [9:28] and grand claims. [9:30] Favor specific and concise responses. [9:32] Make every word earn its place. And [9:35] never use AI slop words like delve, [9:37] foster, leverage, and so on and so [9:39] forth. So, basically adding this stuff [9:41] to the custom instructions helps it [9:43] avoid trying to suck up to you too much [9:45] or using too much AI slop words, which [9:47] is really annoying. [9:49] But again, overall for this use case, I [9:50] think ChatGPT is the winner due to a [9:53] less judgmental personality and also [9:56] generous rate limits at a $20 per month [9:58] plan. [9:59] All right, let's talk about writing and [10:01] editing next. And this is another big [10:03] reversal and is closely tied with [10:05] personality. [10:07] I think Claude's writing skills has [10:09] really devolved over the past few [10:10] months. It's picked up a little bit too [10:12] many Claudisms, like the uncomfortable [10:14] reality, this is X not Y. This just [10:17] drives me crazy. [10:18] And ChatGPT in contrast sticks to my [10:20] preferred writing style a lot better. So [10:22] again, I'm going to have to give this to [10:24] ChatGPT. I also want to give you a quick [10:26] tip on how to make ChatGPT or whatever [10:29] model you use write more like you. All [10:31] right, so here's ChatGPT again. And you [10:33] can see that I'm using a newsletter [10:35] skill to edit this blog post. So, what [10:37] is this newsletter skill? Let's take a [10:38] look. [10:39] So, this is the actual skill file and [10:42] basically what it does is it has a [10:44] writing style guide for my voice and [10:46] tone based on a bunch of writing [10:49] examples that I gave AI. So, it's got [10:51] all this stuff. It's got specific [10:53] directions for how to stick to my [10:55] writing style. [10:56] And I use this skill all the time, not [10:58] necessarily to write from scratch, but [11:00] to edit some first draft that I already [11:01] have. So, here it is. It's edit of my [11:04] first draft. Let me close this. [11:06] And here's the draft and at the very [11:08] bottom it's going to say what it changed [11:11] and also it's going to give a score for [11:14] itself on voice, hook, substance, and [11:16] payoff. And for some reason it's always [11:19] going to give a score around 36 to 38. [11:21] So, maybe it doesn't work properly, but [11:23] I do have evals in place to check [11:25] whether it followed the directions of my [11:28] skill. [11:29] And of course, the The skill also runs [11:31] my no AI slop skill, which has over [11:34] 5,000 GitHub stars now. You can go check [11:36] it out for free on GitHub. And it [11:38] basically just removes all these AI slop [11:40] patterns from ChatGPT or the worst [11:43] offender, which is unfortunately Claude. [11:46] Now, by the way, if you want to have [11:47] access to my newsletter skill and a [11:49] dozen other handcrafted skills from me, [11:51] check out behindthecraft.com [11:53] and become a member. It also comes with [11:56] a full course on how to set up your [11:58] personal AI system and a full year of [12:01] access to me in our private community. [12:03] So, definitely check it out. [12:05] All right. Now, let's move to our next [12:07] use case, which is planning. [12:10] So, whenever I'm trying to make a [12:12] difficult decision, like what paid offer [12:15] should I launch, what project should I [12:17] stop working on, and what should I focus [12:19] on for next 3 months, [12:20] I like to consult Claude Fable first. [12:23] And that's because I think Claude, at [12:25] least as I'm making this video, is [12:27] currently the smartest and wisest model [12:30] in the market. [12:31] So, let me show you what I mean by [12:33] asking ChatGPT and Claude to read my [12:36] planning doc and without sharing [12:38] sensitive information, tell me what I [12:40] should focus on for the next 3 months. [12:42] So, here's ChatGPT. It's saying that I [12:43] should build out my personal AI course, [12:46] spend a quarter improving activation [12:48] renewals, [12:49] and also do other things. In comparison, [12:52] here is Claude Fable's response. [12:54] So, it went through the plan doc and it [12:56] says, "Make the membership launch work." [12:59] Yep, that's basically what ChatGPT said. [13:01] But also, it said this, which ChatGPT [13:03] kind of missed. It says, "Close the [13:05] video ops gap this week." [13:07] And it's telling me to do it cheaply. [13:09] So, basically, I have an admin that's [13:11] been helping me handle making clips and [13:14] podcast post-production for my videos, [13:17] but he's not working with me anymore [13:19] after August 19th. So, I should probably [13:21] go find another ops or VA hire to [13:24] replace him. Cuz I have a problem where [13:26] I tend to do too much work myself or [13:27] with AI instead of delegating to other [13:30] people. [13:30] And Fable was able to find this, but [13:33] ChatGPT was not. So, I think maybe this [13:36] is just one example, but I just feel [13:37] like Fable is a little bit smarter, a [13:39] little bit more thought through when [13:41] it's giving advice for complicated [13:43] decisions like this. [13:45] All right. [13:47] So, I I think for planning, I'm going to [13:49] have to give it to Claude still. [13:52] At least until maybe ChatGPT Astro, the [13:54] new model, comes out. So, now let's talk [13:56] about coding. [13:58] Now, I want to copy here that [14:00] I've shipped a few open-source projects [14:02] and apps for myself, but I'm not a [14:04] software engineer. And in fact, when I [14:06] build stuff with AI, I don't actually [14:08] look at the code at all. And this might [14:10] be blasphemy if you're an engineer. [14:12] So, for my projects, what I like to do [14:14] is I like to plan with Fable, cuz again, [14:16] it's the best at kind of long-term [14:18] thinking. [14:19] And I like to code with ChatGPT. I think [14:22] GPT So, is a model that has a dog in it. [14:26] It'll run tests with browser use, fix [14:28] bugs, and won't stop until the feature [14:30] actually works. Let me show you a [14:32] typical conversation that I have with AI [14:35] on coding. So, here it is building my [14:37] human review skill, and I'm telling you [14:39] there's a few issues. And one really [14:41] good thing about the ChatGPT harness and [14:42] So, is that it's really good at browser [14:45] use. So, it's able to just like use a [14:47] browser and test the product itself and [14:49] make it better without you even being [14:51] there. But basically, yeah, it's just [14:52] like a lot of back and forth [14:53] conversation telling him to fix bugs, [14:55] make improvements, [14:56] and so eventually we get to a point [14:58] where the product actually works. And I [14:59] also think ChatGPT is better at managing [15:02] these long-running conversations and [15:04] context. With Claude, I have to tell it [15:06] to constantly compact context, but [15:08] ChatGPT is able to just like have these [15:09] long-running conversations without [15:12] getting confused. [15:13] So, at least for me, ChatGPT wins on [15:16] coding. But there is an important caveat [15:18] that I want to say, which is that my [15:20] friend Kun, who actually is an engineer, [15:23] he was an L8 engineer, [15:25] believes that GPT-4 so tends to [15:28] over-engineer products too much. And he [15:30] knows this because he reads the code. [15:32] So, basically, he thinks both GPT-4 and [15:34] Opus over-engineer features, and he [15:37] actually recommends using Grok [15:40] to make changes faster and also to make [15:42] more specific changes without making [15:44] like 500,000 line changes. [15:46] So, Kun actually prefers Grok to GPT, [15:49] which is pretty interesting. [15:51] But, you know, since this is my vid [15:52] video, [15:53] I'm going to have to give the award to [15:56] GPT for coding, and maybe you should [15:58] talk to a professional engineer who [15:59] might have a different opinion. [16:02] All right. Now, let's talk about what [16:03] may be the most underrated category on [16:05] this chart, which is browser and [16:08] computer use. The fact is, there's a lot [16:10] of knowledge work that still lives [16:12] inside websites and apps that either [16:14] don't have an API or you have to pay to [16:17] use the API. So, perfect example is [16:19] this. [16:20] Filling out documents, you know? There's [16:22] so many government websites you have to [16:24] navigate, you got to upload a bunch of [16:25] files, it just takes a lot of work. So, [16:27] this person, who immediately works at [16:30] OpenAI, basically just used ChatGPT's [16:31] browser use to prepare a whole [16:34] immigration package in minutes by [16:36] scraping the last 7 years of taxes, bank [16:38] statements, and documents. Now, I [16:40] prepared this immigration package for [16:41] myself and other people before, and I'm [16:44] telling you, this stuff takes hours and [16:46] maybe days. So, the fact that he was [16:48] able to get GPT to do it in a few [16:50] minutes is extremely impressive. [16:52] Of course, you don't want to mess around [16:53] with the US government, so you want to [16:55] actually review everything that GPT does [16:57] to make sure that it's doing it well. [16:59] But, overall, I think this is just a [17:00] massive time-saver. [17:02] Okay. So, to illustrate the power of [17:04] browser use, let me kind of show you the [17:06] social thread that I used to make all my [17:09] social media posts. So, for example, [17:11] recently, I tweeted about my Marie Kondo [17:13] Grok bot and attached the image to it, [17:16] and I basically just use my social post [17:17] skill [17:18] and I said on X you should quote tweet [17:20] the original announcement post for the [17:22] Grok Bot tutorial. [17:24] And basically it drafted this. [17:26] So, here we are. [17:27] And then I made some edits directly and [17:29] I told her to post on all three [17:31] platforms and post on Substack, too, the [17:33] same post. So, it went off and just [17:35] posted everything for me. And like if I [17:37] had to do this normally, I have to go to [17:39] each individual tool or even I have to [17:41] use Typefully, it just takes a lot more [17:43] work and clicking around than just ask [17:45] AI to do it for me, right? And it also [17:47] understands the nuances. Like for [17:49] example, it understands that you should [17:51] tag @GrokBot, which is the handle for [17:53] Grok Bot, on X, but you should not tag [17:57] it on LinkedIn and Threads because that [17:58] handle doesn't actually exist. So, [18:00] that's just a very simple example of [18:02] browser use, but overall it's just [18:03] incredible for any kind of annoying form [18:06] you have to fill online, any kind of [18:08] crappy government or health care website [18:10] that you have to use, or even don't tell [18:12] anyone, but you can even use browser use [18:14] to watch corporate training videos [18:16] without you having to be there yourself. [18:19] Okay? [18:21] So, that's browser use and as you can [18:22] imagine, I'm going to have to give the [18:24] award to ChatGPT. I think the other [18:27] guys, Claude, I think it's not as good [18:29] as ChatGPT, although I think it does [18:30] have browser use. [18:32] Grok Grok Bot is actually pretty good [18:34] for a dedicated personal computer, but I [18:36] think its browser use is also not as [18:38] good. And I think Gemini is still work [18:40] in progress. In fact, I think browser [18:42] use is probably the main reason why I [18:44] love the ChatGPT desktop app so much. [18:47] Now, another really important reason why [18:50] I love ChatGPT is voice chat. [18:53] And this is another category where I [18:55] think ChatGPT is clearly ahead. [18:58] So, basically what I do is I use a live [19:00] voice thread in the desktop app to [19:02] orchestrate my other agents and threads. [19:05] So, let me give you a quick demo of [19:06] this. In the ChatGPT desktop app, you [19:08] can start a voice conversation right [19:10] here. [19:11] And let me show you how it works. I [19:13] don't think you can actually hear the [19:14] voice, but you can see the text as it [19:17] generates. [19:18] Hey, can you start a newsletter thread [19:20] in personal OS to edit my recent post, [19:23] follow up on my chief of staff thread on [19:25] any open items, and see what work is [19:27] left for the video intro skill thread. [19:31] All right, so you can see here that I [19:32] basically asked it to orchestrate three [19:34] other threads that I have. And with [19:35] voice, you kind of do have to remember [19:37] names of your other threads, which is [19:38] why everything here is like named. [19:41] It's going to take a look and it's going [19:43] to give instructions to the other [19:45] threads to start working. [19:47] All right, so it created a new edit [19:50] newsletter post thread, and it gave [19:52] directions to this new thread to start [19:54] working. And it's also orchestrating [19:55] some of the other threads. [19:57] Okay, so you can imagine how powerful [19:58] this is. Like if you want to be ultra [20:00] lazy, you can just be talking to you [20:03] ChatGPT voice while out on a walk, or [20:06] maybe lying in bed, and just have it [20:08] orchestrate the other threads to get [20:10] work done, which is what I do usually in [20:12] the afternoon after having a big lunch. [20:15] All right, so when it comes to voice [20:17] chat, I think ChatGPT is the winner by [20:19] far. [20:20] Now let's finish off by talking about [20:22] image gen, video gen, and personal [20:24] agents. [20:25] So, image generation. [20:27] To find a winner for this, I tested [20:28] making an infographic by just copying my [20:30] entire newsletter post here on how we [20:32] use computers is about to change [20:34] forever. I just copied the whole thing. [20:37] And I pasted this whole thing into [20:39] ChatGPT, Grok, and Gemini because Claude [20:42] doesn't have image gen to see what kind [20:44] of infographic it comes out with. [20:46] So, let's take a look at some of the [20:48] less good options first. [20:50] All right, so this is Grok. So again, I [20:51] just said make an infographic for this, [20:53] keep the text light and clean, and [20:55] pasting the whole post. [20:56] And here is Grok's infographic. [20:59] Yeah, this leaves a lot to be desired. [21:01] It's using like clip art style art, [21:03] which is pretty terrible. And overall, [21:05] it just looks really, really basic. Now, [21:07] let's take a look at Gemini's version. [21:09] Okay, so here is Gemini's infographic. [21:12] Again, the same prompt. And this looks a [21:14] little bit better. This is like uh a [21:15] little bit more detailed. The diagrams [21:17] actually make sense. The text also makes [21:19] sense. But again, I specifically told it [21:21] to keep the text light, and it's [21:24] overcrowding on the text. It also has [21:26] some weird stuff that it added, like [21:28] nano banana, which is kind of weird. But [21:30] I guess that it added that because I [21:32] told it to make it using a nano banana. [21:34] So, maybe that's my fault. But I think [21:36] this is definitely better than Grok, but [21:37] maybe not as good. [21:39] Next, I asked ChatGPT to make the same [21:42] infographic. And this is changed a [21:44] little bit because ChatGPT actually has [21:46] full context of my past conversations. [21:48] And I've had a lot of conversations with [21:50] it about my brand guidelines, and also [21:53] this post by itself. But let's take a [21:55] look at the infographic anyway. [21:57] So, this is the infographic. Yeah, I [21:59] think this looks pretty good. It looks a [22:00] lot better. I think the images are still [22:02] a little bit too big, but at least it [22:04] has my brand style guidelines with the [22:06] circle, and my colors, and the [22:08] illustration style that I like. [22:11] All right. Overall, I think I have to [22:12] give the winner here again to ChatGPT. [22:15] It seems to be better at following my [22:17] instructions than the other two options. [22:19] But I think a Gemini's nano banana is [22:21] also quite good and comparable if you [22:24] give it the right instructions to reduce [22:25] the amount of text that it shows in the [22:27] image. So, let's give this to ChatGPT. [22:31] And now, let's talk about video [22:32] generation, which is always super fun. [22:35] Only Grok and Gemini can generate [22:37] videos. So, let's compare the two. [22:40] All right. [22:41] You know, one of my favorite family [22:43] activities with my kids is to generate [22:45] crazy Japanese video commercials [22:48] featuring our images. So, here I'm [22:50] asking it to generate a picture of [22:52] myself, and generate a crazy Japanese [22:54] video commercial featuring me [22:56] advertising some delicious ramen, and [22:58] make it as crazy, and make the voiceover [23:00] in Japanese, too. All right. Let's take [23:03] a look at Gemini's creation. [23:05] And here we are. I don't think you [23:06] actually hear a sound. Let me go ahead [23:08] and play it. All right. So, you can see [23:09] here [23:10] it's even more fun with the sound on. [23:12] But, this is already pretty amazing, [23:13] right? This is pretty good stuff. [23:15] I don't think I'm ever going to share [23:16] this video publicly to save face. But, I [23:18] think this definitely got a laugh out of [23:19] me and my kids. [23:21] Now, let's take a look at Grok Imagine's [23:23] crazy Japanese commercial. [23:24] All right. [23:26] This is uh [23:27] This is a little bit too much. It's [23:28] exaggerating in my face a little bit too [23:30] much. And I think it's actually pretty [23:32] damn scary. Like, the fact that I'm [23:33] jumping and going crazy about this. I [23:35] don't know what's going on here. So, [23:36] yeah. Overall, based on this like single [23:38] example, I'm going to have to give video [23:39] generation to Gemini. I do think Grok [23:42] probably has less restrictions in place [23:44] for the types of videos you can [23:45] generate. And of course, there's also [23:47] like Chinese tools like SeaArt that have [23:48] no restrictions whatsoever. [23:50] But, overall for like family videos, [23:52] crazy Japanese commercials, I think [23:54] Gemini is the winner here. [23:56] Last but not least, I want to talk about [23:58] personal agents. [24:00] And by this I mean like an AI that can [24:01] help you get work done across your [24:03] favorite apps. I think the personal [24:05] agent race is actually less about the [24:07] model and more about the harness or the [24:10] tool. So, the ChatGPT desktop app that I [24:12] demoed before is still my default [24:14] because it's best in class at browser [24:17] use, voice, and mobile access features. [24:20] I can get a little bit overwhelmed with [24:21] all these threads over here and also [24:23] like the different modes that you see [24:25] here between Codex and ChatGPT and chat [24:28] and work. [24:29] But, overall I think it's just the most [24:30] robust and flexible harness. [24:33] Now, Claude Code is also a pretty great [24:35] harness. But, I think it's really [24:37] playing catch-up still with ChatGPT on [24:39] browser use, voice, and mobile. And Grok [24:42] bot is kind of the new player. And it [24:44] has much cleaner UX than ChatGPT. And I [24:46] also love how it gives all the Grok [24:48] agents a 24/7 dedicated cloud computer [24:52] where they can all live on. [24:53] And this is overall pretty well [24:55] designed. But, I think it's a little bit [24:57] less flexible. For example, for each of [24:59] these agents, I actually want to talk to [25:02] them across multiple chat threads [25:04] instead of just one. And right now, it [25:06] feels a little bit too restrictive. [25:08] And actually, many people don't know [25:09] this, but Gemini also has a personal [25:11] agent called Spark. [25:13] And it's also pretty interesting and [25:15] pretty well designed. I think it [25:18] integrates well with Google Workspace [25:19] and recently added third-party [25:20] integrations. [25:22] But, it doesn't have all these crazy [25:23] plugins [25:25] that ChatGPT already supports. [25:28] All right. So, overall, I'm going to [25:29] give this last category to ChatGPT as [25:32] well. [25:33] But, I think if Grokbot gets more robust [25:36] and flexible, I might actually switch [25:38] over to Grok at some point. [25:40] And for what it's worth, I think this is [25:41] the probably the most important use case [25:43] for AI productivity, even more important [25:45] than coding. And whoever wins the [25:47] personal agent race, I think will [25:49] capture the lion's share of the consumer [25:52] attention of AI. [25:54] All right. So, there you have it. This [25:55] is my complete table with my [25:57] preferences. As you can see, ChatGPT is [25:59] the clear winner for me right now. And [26:02] at the pace that the ChatGPT and Code [26:04] Assistant is shipping, [26:05] I'm not sure they're ever going to give [26:06] up the lead. [26:07] But, that being said, I think Claude is [26:09] still really great for design and [26:11] planning. Gemini is great for video [26:13] generation and also is a little bit more [26:15] simple to use than ChatGPT. And Grok is [26:18] now a great contender, especially with [26:21] Grokbot for personal agents. [26:23] But, for most of you, I think you can [26:25] get by with just $20 a month using [26:28] ChatGPT and getting access to all these [26:30] features is a really good deal. [26:32] All right. If you enjoyed this video, [26:34] please check out behindthecraft.com [26:36] for my full course on how to set up your [26:38] personal AI system to save time and get [26:41] more work done. [26:42] Also, like and subscribe to this video [26:44] and I'll see you next time.