# How to Actually Code with AI (ft. Matt Pocock) ![[images/How-to-Actually-Code-with-AI-Matt-Pocock.webp]] **Creator:** NetworkChuck (Chuck Keith) · **Guest:** Matt Pocock · **Published:** 2026-11-20 · **Length:** 2:02:36 · [Watch on YouTube](https://www.youtube.com/watch?v=mU8WKII1gpk) > Caption source: YouTube English auto captions (`en-orig`, about 24,800 words) downloaded with yt-dlp (android_vr client) on the first try. Auto captions garble names: "Matt PCO" = Matt Pocock, "Johnow"/"John Asterout" = John Ousterhout, "Versel" = Vercel, "Netlefi" = Netlify, "S castle" = Sandcastle, "Everlight"/"Everite" = Evalite (both identified from the GitHub pages shown on screen), "Grockbot" = Grok Bot, "Dex Hy" = Dex Horthy, "Kemp Beck" = Kent Beck, "Carpathy" = Karpathy. Skill names, commands and file names were checked against frames and against Matt's public `mattpocock/skills` README and aihero.dev pages. The 16 headings follow the creator's own chapters. Added from the user's "Summarize" playlist. ## Outline ### 1. Do you still need to learn a programming language? ([0:00](https://www.youtube.com/watch?v=mU8WKII1gpk&t=0s)) - Cold open, taken from later in the interview: Pocock says nobody needs to learn a specific programming language right now, and that "coding has always been easier than people think." Chuck: "I do not have a software development background, yet I can now do it all... But I can feel the limitations." ### 2. Who Matt Pocock is ([0:54](https://www.youtube.com/watch?v=mU8WKII1gpk&t=54s)) ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t0135.webp]] *2:15: Matt Pocock's lower third, "Total TypeScript · AI Hero"* - Chuck bought Pocock's AI course and found it "awesome" but challenging, which is why he wanted the conversation. Pocock has just shipped a new AI crash course and skills v1.3, and has a baby due in late November. - Pocock: "right now I'm the AI guy... the skills guy." Before that he was "the TypeScript person": open-source work on **XState** (state machines and statecharts in TypeScript) taught him a lot about types, which led to his course **Total TypeScript** and full-time teaching. He was skeptical of AI at first and "has risen through" it. He now writes down processes that work and ships them as markdown files (skills). - Chuck's framing: AI has handed people "a magic wand" but they "need to go to Hogwarts to learn how to use this thing." Ask your agent how to share an app and "they spin up Vercel or they do something random and it's their opinion and it's not good." ### 3. Tactical vs strategic programming, and why vibe coding wrecks code ([5:03](https://www.youtube.com/watch?v=mU8WKII1gpk&t=303s)) ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t0315.webp]] *5:15: the episode's "TACTICAL: AI is great at this" card* - Pocock rejects the guru framing ("I've got a bald head... I live near Oxford. I've got glasses. That doesn't make me smart"). - **John Ousterhout's *A Philosophy of Software Design*** separates **tactical** programming (the sergeant winning today's battle) from **strategic** programming (the general winning the war). "AI is really good at tactical programming and terrible at strategic programming." Benchmarks judge models on a single task, so they are trained to be 100% tactical, and strategic mistakes "don't usually come back at you until like 9 months down the line." - **Vibe coding** has suffered "semantic diffusion." He uses the original meaning: coding without looking at or thinking about the code. Done exclusively, it never works, because the code is the agent's environment and it slowly gets worse, like the "this is fine" dog sitting in a burning house. "You need someone to keep the house in order and that person is probably at the moment you." - Not using AI is "ridiculous": most of the old job was tactical and "now you can pay less than minimum wage for someone else to do it." He is happy that "tactical programming is dead." - Can a non-engineer become a software engineer? "Software engineering has always been easier than people think it is." Strategic programming "is just caring a bit more than the person next to you." ("You just got to give a [expletive].") - Sponsor (11:18 to 12:35): **ThreatLocker's Zero Trust World**, Feb 17 to 19, 2027, with hands-on hacking labs, talks and CE credits. Code ZTWNETWORKCHUCK27 for $200 off. ### 4. Live demo: Grill with Docs ([12:35](https://www.youtube.com/watch?v=mU8WKII1gpk&t=755s)) ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t0870.webp]] *14:30: the glossary of Pocock's `course-video-manager` repo, the domain language the agent reads before it grills him* - His skills "have evolved into a process," and strategic thinking is "blended into every stage." The demo repo is his **course video manager**: a video editor plus AI writer where he produces all his content. - **Start with language.** This comes from **domain-driven design**. He holds up **Eric Evans' *Domain-Driven Design*** ("a bit of a tome") and says the first chapters on language are "so so cool." - **`/grill-with-docs`** quizzes ("grills") you about what you are about to build. It is a richer version of his most famous skill, **`/grill-me`**. He gives it a deliberately vague prompt: *"I need to improve the teleprompter part of my video editor. I think it's just gotten a little bit cluttered and I think there's some stuff I no longer use in there. Can you help me out?"* - The agent explores the codebase with a sub-agent and reads the **glossary**: courses, sections, course versions, lessons, drafts, pending versions. Because the agent already speaks the app's language, even a verbose model (he names Opus 5) stays concise. - He dictates with **Wispr Flow** ("if you're not dictating, you're just going way too slow"). Chuck uses it too. - The agent asks what the teleprompter reads from, what is on screen while filming, and what clutter he has noticed. Pocock: "we're sort of slowly getting aligned." There is always a communication gap: the agent is like a hired developer dropped in "like the guy from *Memento*... with no memory whatsoever." Forcing it to ask questions gets you "the result that you want instead of the result that you sort of hope for." - Chuck's example: an agent added an unrequested feature to his interview tool. Pocock has been a teacher since he was 21 (14 years) and still gets complacent when prompting: "it's the human that's in the wrong." ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t1260.webp]] *21:00: `GLOSSARY.md` entries such as "Published Version" and "Submit", with the terms to avoid* - How the skills fit together: `/grill-me` now just says "call the skill tool with grilling." `/grill-with-docs` calls **grilling** plus **domain modeling**: the glossary plus **ADRs** (architectural decision records), short markdown files about hard-to-reverse decisions. "Code is... only the present. It doesn't tell you about the past," so without ADRs the agent repeats the same mistakes. His repo has about 31 ADRs. - Every skill tells the agent to use the glossary, so code, database and UI share one vocabulary. His counterexample: the front end says "reversal," the email says "pending chargeback," the database says "refund." - You do not have to write the glossary up front. Just start grilling: the agent mints terms as you go and offers to update the code to match, "peeling like a little window of sanity in your vibe coded slop base." - Where to get it: **aihero.dev/skills**. "For your next feature, just use grill with docs." Both men admit to Twitter/X addiction; Chuck uninstalled it and is coding a tool to stay updated without it. ### 5. The smart zone and the dumb zone: write a spec first ([25:18](https://www.youtube.com/watch?v=mU8WKII1gpk&t=1518s)) ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t1650.webp]] *27:30: aihero.dev's page for the `/implement` skill ("Build a finished spec into code, test-first")* - After grilling there are two routes. If the job fits in the same context window, go straight to **`/implement`**. - "I'm not really talking about the 1 million token context window that Opus has. I consider there to be a **smart zone** and a **dumb zone**." As tokens are added, the relationships between them grow quadratically and attention degrades. "I start getting nervous around **150k tokens**." Small contexts are cheaper, faster (latency grows with tokens) and smarter. The exact number matters less than "the instinct of **context paranoia**." - If the work will exceed about 150k tokens, turn the grilling conversation into a **spec** (a markdown file or a GitHub issue), then turn the spec into **tickets**. Each ticket is one session, followed by an adversarial **code review** against the spec. ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t1680.webp]] *28:00: the `/to-tickets` skill page ("Split a spec into small tickets an agent can build")* - Context tactics: big codebases do not use more context in themselves; *hard-to-navigate* ones do. (He forgets his "four things" mid-answer. Chuck: "Oh no, you compacted your context.") - **Effort level:** he uses **Opus 5.5 on medium** for nearly everything and is considering low. He thinks there is "an epidemic of people using too high an effort": on exploration, extra effort just burns tokens and pushes you into the dumb zone sooner. He goes higher only for web research, performance testing and hard debugging. Chuck runs extra-high and says he will try medium. - He has one Anthropic subscription ("200x") and does not min-max between models: "instead of waiting for the next model to fix all your problems, why not just fix all your problems?" Chuck has three Anthropic Max plans plus Codex plans. ### 6. Sub agents in worktrees, and code review ([34:35](https://www.youtube.com/watch?v=mU8WKII1gpk&t=2075s)) ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t2100.webp]] *35:00: the `/implement-spec` skill page ("Build a whole spec end to end, with parallel subagents")* - **`/implement-spec`**: in a new session, pass in the spec. It implements it "with parallel sub agents doing it in work trees, merging all back together until you get one mega PR" to review. - He credits the smart zone / dumb zone idea to **Dex Horthy** ("who you should have on"). Tickets should be about 150k tokens each, and are tracked as **GitHub sub-issues** under each spec. - **`/code-review`** runs two reviews: a **spec review** (did it build the right thing, over or under scope?) and a **standards review** against a `CODING_STANDARDS.md` in the repo. His uses **Effect** ("a very cool TypeScript library") and includes rules like "every `any` in TypeScript is a leak." ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t2310.webp]] *38:30: Pocock's `CODING_STANDARDS.md`* - Chuck admits he lets the agent write its own standards and never reads them. Pocock: "I would probably describe that as vibe coding... I want you to be good, but I don't care how you're good." (Both joke about "ultrathink... or I'll fire you.") - **Human review** is only for important work. His **`/pr`** skill writes a PR body with a visual summary, evidence it works (a screenshot or execution output) and the **merge danger**: a **one-way door** (can't be undone: data loss, an email to 100,000 people) or a **two-way door**, plus the **blast radius**. A two-way door with a small blast radius he may "just test in production." A one-way door with a big blast radius he reviews by "trying to break it." ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t2550.webp]] *42:30: the PR template: summary, evidence, merge danger (one-way or two-way door, plus blast radius)* - Whatever he writes in review goes back into the environment: "how can I improve my environment so I never have to write this comment again?" The **`/retro`** skill looks back at a session or PR and suggests improvements. - Then merge ("okay guys merge all these five PRs... figure out the right order"). The loop: **grill, then spec and tickets (or straight to implement), then code review, then retro.** ### 7. The trust ladder: tests, types and feedback loops ([44:52](https://www.youtube.com/watch?v=mU8WKII1gpk&t=2692s)) - **Lauren Tan** ("Potato"; the captions say she works at "Grockbot") is "famous now for shipping like 2,500 PRs a month." Her idea is a **trust ladder**: if you can't trust your agents you'll never run more than about four at once, because you'll watch everything they do. Better environments let you trust hundreds. - The classics agree that a good codebase is one that is **easy to change** without breaking things. That needs lots of **feedback loops** so the agent can grade its own work. The one place he *does* want high effort is **verification**: clicking through the real app and running performance traces. - Does he trust his agents? Yes, because his software is personal, not medical or financial. On banking apps he would scrutinize one-way doors and spend heavily on security verification. ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t2940.webp]] *49:00: the free AI Coding Dictionary, open at "Automated check"* - His free **AI Coding Dictionary** is a graph of every term he uses (model providers, context windows...), with a GitHub repo you can point agents at. - **Automated checks:** strong types (TypeScript, not JavaScript), fast runtime **tests**, and **linters** (check functions, imports and bad patterns). Get the most out of them, e.g. a standards rule banning `any` because "it's actually a cheat." ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t3030.webp]] *50:30: the rule "Every `any` is a leak" in `CODING_STANDARDS.md`* - To add checks, just tell the agent: "I want to constrain my application more. I want you to find harsh lint rules." `/retro` also tries to turn PR comments into deterministic checks, ratcheting up until agents are "in their straight jacket" and "literally cannot make a mistake." Then you work on keeping the checks fast. ### 8. The starter stack and deep modules ([52:31](https://www.youtube.com/watch?v=mU8WKII1gpk&t=3151s)) - For an "average Joe" app idea (cheap coffee and tea finder), his free stack is **GitHub, VS Code, Git and Claude Code**, or any harness that supports skills ("they've all copied each other"). - `/to-spec` asks what shape of modules you want. He rejects the "**dark factory**" (agents working with the lights off) in favor of "little dimmer switches" for rooms you rarely enter. - **Tests are hard.** Bad tests repeat the implementation (e.g. a test that a tweet limit constant equals 240). Good tests treat a big **module** as hidden and test through a narrow entry point, so you can change its insides without touching tests. These are Ousterhout's **deep modules**: "you can actually sort of turn the lights off inside that module." ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t3420.webp]] *57:00: the `/improve-codebase-architecture` skill page* - **`/improve-codebase-architecture`** produces a visual report of problems and proposed improvements, which you can turn into a spec and tickets. It means unit tests that "take out huge chunks of the application and put it all under test at once." Vibe-coded codebases inspire "terror." Deep modules that use glossary terms let you think in chunks (auth, transcripts, the video editor). - Sponsor (1:00:00 to 1:01:01): **Coveron** identity and fraud protection (dark-web monitoring, up to $2 million identity-theft coverage, 71% off Black Friday deal). ### 9. Improve Code Base Architecture, live ([1:01:01](https://www.youtube.com/watch?v=mU8WKII1gpk&t=3661s)) - He runs `claude` and then `/improve-codebase-architecture`. It pulls in the **codebase-design** skill, finds hot spots from recent commits and returns an **HTML report**. - "There's no one perfect codebase." A **cron job** on each repo runs the skill daily, picks the top suggestion and opens a PR. These are two-way doors, so he **samples** about one in five and occasionally runs `/retro` on the last five. - He just tore down a custom Discord bot ("my own Hermes") he texted from his phone: "phones are cigarettes... phones are poison." Now he uses Claude Code plus **GitHub Actions** for scheduled jobs (`claude` inside the action opens PRs, free for open-source repos). - Chuck texts his agent over iMessage and loves it, but just deleted X, TikTok and YouTube from his phone. Pocock also removed the browser and email from his Android phone, and "nothing rushed in to fill that gap." ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t4050.webp]] *1:07:30: the HTML report, showing duplicated course-copy code before (shallow modules) and after (one deep module)* - The report: "submit and duplicate course contain 150 near identical lines." It proposes replacing several shallow modules with one deep, testable module. Chuck says most viewers would need to ask the AI what half of it means. ### 10. What you still need to learn ([1:08:15](https://www.youtube.com/watch?v=mU8WKII1gpk&t=4095s)) ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t4125.webp]] *1:08:45: the `/teach` skill page with its install line `npx skills@latest add mattpocock/skills --skill=teach`* - "There's a skill for that." **`/teach`** (written "bored on the tube one day") turns a folder into a teaching workspace: it tracks what you know, quizzes you and builds HTML lessons. He tested it by learning a **Rubik's cube**. Pass it a report you don't understand: "Teach me about this." - He holds up **Michael Feathers' *Working Effectively with Legacy Code***. His skills use terms from well-known books, e.g. **locality**: how likely a change is to stay in one place. - Will models make this unnecessary? You need "enough strategic programming to be dangerous." Engineers produce better products because they know the words: "fix the hydration error" or "fix the layout shift on the homepage" instead of describing things jumping around. - On "**the bitter lesson**": "I'm not a pundit... I'm just in rural Oxfordshire." Some harness may make agents care, but for now this is "the state-of-the-art way." ### 11. Where to host your app, and deploying safely ([1:14:11](https://www.youtube.com/watch?v=mU8WKII1gpk&t=4451s)) - No skills needed. DevOps moved from degree-level work to managed products (Google Cloud, AWS, serverless, and platforms like **Vercel**, where he used to work, and **Netlify**). Just decide where data lives: "probably just use like a **Postgres** database" from a host such as **PlanetScale** or **Supabase**, then deploy to Netlify or Vercel. Add regular security reviews. - Chuck, "very much a home labber" on **Proxmox**, has his agent spin up VMs, IPs and domains. Pocock is on Vercel: "push to your GitHub repo and it deploys." **GitHub Pages** and **Cloudflare** also work. Prompt: "make this deploy as good as possible... DDoS attacks... give me the basic level of what I need." - Sponsor (1:18:28 to 1:20:01): **Bitdefender RealCheck**, a free deepfake-check app. Chuck saw a fake Elon Musk crypto livestream with about 30,000 viewers while looking for the Roadster announcement. - Security: "I'm the wrong person to ask." His projects are personal software and a platform team handles his site. Recommendation: add a **security review** to code review (many harnesses have one built in). - **Skill evals** are "incredibly hard and incredibly expensive." His evals are his users (he says "millions") and an active Discord. A daily scheduled agent reproduces reported issues in a dummy repo and opens PRs, filtering out "crap issues." - Deployment is **CI/CD**: GitHub pushes to Vercel, which rebuilds. "If you have a choice between determinism and non-determinism, why would you ever pick non-determinism?" The same goes for retro: prefer a lint rule over a review rule. Chuck self-hosts **Gitea** and uses its Actions. - Not a SaaS guy: teaching is a lifestyle choice so he can be home with his young son (a second child is on the way). He has turned down job offers and is "basically unhirable." ### 12. Open source in the age of slop PRs ([1:26:38](https://www.youtube.com/watch?v=mU8WKII1gpk&t=5198s)) ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t5250.webp]] *1:27:30: the Sandcastle README, "Local Coding Agent Orchestration in TypeScript"* - His open source: **Sandcastle** (agent orchestration in TypeScript, running any agent in Docker, Podman or Vercel sandboxes) and **Evalite** (a local visual eval framework that has "fallen by the wayside"). The bulk of his effort goes into Sandcastle and the video editor. - Contributing has changed in the 18 months since it was easy to "ping up a slop PR." Popular projects are closing PRs: **Hono** ("one of the most popular kind of like server frameworks in JavaScript") now only takes PRs from collaborators, and he closed PRs on his skills repo ("nearly up to 300,000 stars"). ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t5400.webp]] *1:30:00: Hono's pull-request list on GitHub* - Contribute **issues with good evidence and reproductions** instead. Or study great repos: "/teach me about the Hono codebase... What are they getting right strategically?" Publishing to **npm** with a changeset manager in GitHub Actions is easy. - Open source always had a **BDFL** (benevolent dictator for life) and has always been "weird... political"; it is "harder to get into than it's ever been." - Is it still a portfolio? "AI has eaten tactical programming, but it can't do strategic programming," so show strategic evidence: infrastructure, a beautiful repo, steering agents. That can make you a "200x agent user." ### 13. Advice for new developers: school, degrees, hiring ([1:34:41](https://www.youtube.com/watch?v=mU8WKII1gpk&t=5681s)) ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t5895.webp]] *1:38:15: the episode's "Matt's Reading List" card* - High-schooler advice (he hasn't job-hunted "since 20 or something"): **demonstrate that you care** about the strategic layer. Use words like deep modules and ship projects such as a personal CRM you make "as beautiful as you can" and deploy. - No specific language needed ("that's tactical stuff"). Buy the classics, use `/teach` on the hard parts. He found Feathers' book (he says it is "from 2003") still wrapped in plastic on his shelf, "couldn't believe what I was reading," and it inspired `/improve-codebase-architecture`. - **His reading list:** *A Philosophy of Software Design* (Ousterhout), *The Pragmatic Programmer* (David Thomas and Andrew Hunt), Michael Feathers' book, *Refactoring* (Martin Fowler) and *Extreme Programming* (Kent Beck). "If you actually do the reading, it puts you way ahead of everybody else." - Strategic programming is balancing competing goals, e.g. 4,000 tests that slow the agent, so you put them in CI and shard them across four nodes. "You're just giving the agent permission to care." - CS degrees: "how effective have they ever been actually?" He has a **drama degree**, was a **voice coach for six years** and taught himself to code around **2015**. To CS students: "Care." - Hiring: there are still bottlenecks. Two people shipping 2,500 PRs a month is 5,000. Products are differentiated by "the amount of attention" given to them (QA, planning, product management and design). "I've said the word care about a thousand times." - His new course: a two-week cohort starting **November 9**, just before his paternity leave (so the interview was recorded before that date). ### 14. Software factories and Matt's AI video editor ([1:44:39](https://www.youtube.com/watch?v=mU8WKII1gpk&t=6279s)) - The course is about improving the agent's environment, ending with a **software factory**. That has its own entry in the AI Coding Dictionary: it removes the human from more points in the pipeline (triggers from new issues or slow-database alerts, agents that decide whether to act, human review only for one-way doors). You also need brakes "so it doesn't just become a slop cannon." He calls it "**loop engineering** by another name." - His course platform was built from scratch with **badass.dev**. ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t6600.webp]] *1:50:00: the public `mattpocock/course-video-manager` repo* - **The video editor**, born from "I hate editing video": he leaves a gap while recording, the clip is cut and transcribed automatically, and it appears in the finished edit. It grew a diagram playground, a teleprompter (he considered an **Elgato** one), AI-written scripts and animatics with Opus 5.5 ("watch the video I'm about to record before I record"), sync to his course platform, version history, linting and AI writing. It is open source at about 2,500 commits, and he takes neither PRs nor issues. - Chuck's pipeline: offload and ingest footage, transcribe, have the **Gemini** video API watch it, auto-edit a clean draft, review it in a home-built Frame.io clone, then export **Adobe Premiere** XML for his editors. - Pocock uses AI scripts only as a guide (commands to copy, commits to check out) and improvises the rest ("I did a drama degree"). - Course planning: every ~3 hours a cron job ingests tweets, mentions, Discord, Slack and Gmail into a private "**Karpathy-style LLM wiki**." He asks it what his audience wants, turns that into learning goals, then goes **off AI**: he prints them on cards and arranges them by hand into lessons (hours per section; courses often have eight sections). Then come scripts, agent dry runs and animatics. Planning a course "can be months." - Pen and paper is "me-paced"; agents are too fast (or too slow: "I'll just check Twitter... half an hour later"). Afterward he dictates a brain dump back with Wispr Flow. ### 15. Agents for your life ([1:56:49](https://www.youtube.com/watch?v=mU8WKII1gpk&t=7009s)) - His audience wiki is the "home base for all of my agents." He has a daily standup with them, dictates to-dos into a **Pixel Watch**, and an agent triages and schedules his day. He has two small desk terminals but is "still trying to find the right interface." - Knowledge work is "much harder than coding" because coding is verifiable and course quality isn't. That is why agents excel at math but "still can't draw hands" or write a convincing tweet. Both want human things kept human: "keep agents out of my board games," no AI art or music, "I get a thousand AI tweets a day and I hate them." - Where to find him: **aihero.dev**, his skills, Discord, Twitter, and YouTube, which "recently crossed 400,000 subscribers." There is a newsletter for the course. ### 16. A prayer for you ([2:01:01](https://www.youtube.com/watch?v=mU8WKII1gpk&t=7261s)) ![[images/How-to-Actually-Code-with-AI-Matt-Pocock-t7270.webp]] *2:01:10: Chuck's closing prayer* - Chuck's customary closing prayer for viewers: peace about AI and jobs, a job for those looking, ideas for those starting companies, and blessing on their families, "in your name, Jesus. Amen." ## Key takeaways - **AI does tactics; you own strategy.** Strategy is mostly *caring* about the agent's environment: language, tests, types, lint rules, module shape. - **Align before coding.** Grill first, and keep a shared glossary and ADRs so the agent speaks your domain. - **Stay in the smart zone.** Keep sessions small (he gets nervous near 150k tokens), use specs, tickets and sub-agents for big work, and don't crank effort up by default. - **Review by merge danger,** not line by line: one-way vs two-way door, plus blast radius. - **Ratchet the environment.** Turn every review comment into a deterministic check where possible, and prefer determinism (CI/CD, linters) over agent judgment. - **Deep modules** make tests meaningful and let you "turn the lights off" inside them. - **Read the classics** and use AI (`/teach`) to digest them. ## Tutorial: Matt Pocock's skills workflow, step by step Assembled from what Pocock shows and says, with install commands from his public README and aihero.dev (fetched 2026-11-20). Skill names follow the current repo, so check the README before relying on them, because the skills change between versions (v1.3 shipped 2026-10-05). ### Step 1: Set up the stack You need Git, a GitHub account, VS Code (optional but recommended) and a coding agent that supports skills (he uses Claude Code; "it doesn't really matter what the agent is"). ### Step 2: Install the skills Pick **one** method per agent. Installing both gives you every skill twice. ```bash # Claude Code: managed plugin that updates itself (from the README) claude plugin install mattpocock-skills@claude-plugins-official ``` ```bash # Any agent, as editable files in your project (aihero.dev / README) npx skills@latest add mattpocock/skills # a single skill, as shown on screen for /teach: npx skills@latest add mattpocock/skills --skill=teach # update later: npx skills@latest update ``` ```bash # Codex codex plugin marketplace add mattpocock/skills codex plugin add mattpocock-skills@mattpocock ``` Then, once per repo, run the setup skill. It asks which issue tracker you use (GitHub, GitLab, local files...), your triage labels, and where to save docs: ```text /setup-matt-pocock-skills ``` ### Step 3: Grill before you build Start every change with a grilling session. Vague prompts are fine, because the point is that the agent interviews *you*. His demo prompt: ```text /grill-with-docs I need to improve the teleprompter part of my video editor. I think it's just gotten a little bit cluttered and I think there's some stuff I no longer use in there. Can you help me out? ``` - `/grill-with-docs` = grilling + domain modeling. It creates and updates `GLOSSARY.md` (shared terms) and ADRs (`docs/adr/`) as you go. You don't need to write the glossary first. - For non-code planning, use `/grill-me`. - Dictating answers (he uses Wispr Flow) makes this much faster. ### Step 4: Choose a route by size - **Fits in one context (well under ~150k tokens):** `/implement`, which builds the agreed work test-first and closes with `/code-review`. - **Bigger:** turn the conversation into a spec, split it, and build it in a fresh session: ```text /to-spec /to-tickets # new session: /implement-spec <link or path to the spec> ``` `/implement-spec` runs implementer sub-agents in parallel (in git worktrees, per Pocock) and merges into one PR to review. Tickets appear as GitHub sub-issues under the spec. Aim for about 150k tokens of work per ticket. ### Step 5: Keep effort modest Pocock runs Opus 5.5 on **medium** for almost everything, including exploration and grilling. Raise effort only for web research, performance work, hard debugging, and **verification** (when the agent checks its own work in a real app). ### Step 6: Review by merge danger - `/code-review` runs a **standards** review against your coding standards file and a **spec** review against the issue, as separate sub-agents. - The `/pr` skill writes a PR body with a visual summary, before/after **evidence**, and **merge danger**: one-way or two-way door, plus blast radius. Human-review the one-way, big-blast-radius changes (try to break them), and spot-check the rest. - Keep a `CODING_STANDARDS.md`. Example rule from his repo: ```markdown ## Types ### Every `any` is a leak An `any` switches the type checker off for every value that flows through it. Write the type you mean: a known shape gets a named `interface`; an unknown one gets `unknown` and is narrowed; one that varies by caller gets a generic. ``` *(Abridged from the on-screen file at 50:30.)* ### Step 7: Retro, and ratchet the guardrails After a session or PR: ```text /retro ``` It suggests environment fixes (navigation, automated checks, coding standards, steering files, tooling), most severe first. Prefer a **deterministic** check (lint rule, type, test, CI job) over a prose rule whenever possible. To tighten up fast, ask: ```text I want to constrain my application more. I want you to find harsh lint rules. ``` ### Step 8: Improve architecture continuously ```text /improve-codebase-architecture ``` It reads `GLOSSARY.md` and ADRs, uses recent commits to find hot spots, and returns an **HTML report** of "deepening opportunities" (shallow modules into deep, testable ones). Pick one and grill it, or send it through spec and tickets. Pocock runs it **daily from a cron job** (Claude inside **GitHub Actions**, free for public repos) and samples about one PR in five. A schedule along these lines would do it, though he did not show the YAML, so treat this as illustrative only: ```yaml # illustrative only; not shown in the video on: schedule: - cron: "0 6 * * *" # once a day jobs: improve-architecture: runs-on: ubuntu-latest steps: - uses: actions/checkout@v4 # run your agent headlessly here with /improve-codebase-architecture # and have it open a PR for the top suggestion ``` ### Step 9: Deploy with CI/CD, not agents Put data in a managed **Postgres** (e.g. PlanetScale or Supabase), connect the GitHub repo to **Vercel** or **Netlify**, and every push to the default branch redeploys automatically. Ask the agent once to harden it: ```text Make this deploy as good as possible. Give me everything I might need in terms of security, including DDoS protection. Just give me the basic level of what I need. ``` Add a security review to your code-review step. ### Step 10: Learn what you don't understand ```text /teach I don't understand this report. Teach me about this. /teach me about the Hono codebase. What are they getting right strategically? ``` `/teach` keeps state in the current folder across sessions, quizzes you and builds HTML lessons. Pair it with the books below. ## Things mentioned ### Books | Title | Author | Mentioned at | | --- | --- | --- | | *A Philosophy of Software Design* (tactical vs strategic programming, deep modules) | John Ousterhout | [5:06](https://www.youtube.com/watch?v=mU8WKII1gpk&t=306s) (again 56:24, 1:38:02) | | *Domain-Driven Design* (held up on camera; the source of his shared-language idea) | Eric Evans | [14:14](https://www.youtube.com/watch?v=mU8WKII1gpk&t=854s) | | *Working Effectively with Legacy Code* (held up on camera; inspired /improve-codebase-architecture) | Michael Feathers | [1:09:22](https://www.youtube.com/watch?v=mU8WKII1gpk&t=4162s) (again 1:36:45) | | *The Pragmatic Programmer* | David Thomas and Andrew Hunt | [1:38:05](https://www.youtube.com/watch?v=mU8WKII1gpk&t=5885s) | | *Refactoring* | Martin Fowler | [1:38:10](https://www.youtube.com/watch?v=mU8WKII1gpk&t=5890s) | | *Extreme Programming* (*Extreme Programming Explained*) | Kent Beck | [1:38:15](https://www.youtube.com/watch?v=mU8WKII1gpk&t=5895s) | *Note:* the on-screen "Matt's Reading List" card (1:38:15) lists only four: Ousterhout, Thomas and Hunt, Feathers and Fowler. The YouTube description's "Books Matt recommends" lists Ousterhout, Evans and Feathers. ### Other works - "The Bitter Lesson" (Rich Sutton's 2019 essay; Chuck first calls it "the bitter pill"), 1:12:51. - *Memento* (Christopher Nolan film), used as a metaphor for a memoryless agent, 17:16. - The "this is fine" burning-house dog meme, 7:44. ### Matt Pocock's skills (mattpocock/skills) `/grill-me`, `/grill-with-docs`, grilling, domain-modeling (`GLOSSARY.md`, ADRs), `/implement`, `/to-spec`, `/to-tickets`, `/implement-spec`, `/code-review`, `/pr`, `/retro`, `/improve-codebase-architecture`, codebase-design, `/teach`, `/setup-matt-pocock-skills`. ### Tools, models and services Claude Code, Anthropic Opus 5 / Opus 5.5 (and "Fable", which Chuck mentions as an alternative), OpenAI Codex, VS Code, Git, GitHub (issues, sub-issues, Actions, Pages), Wispr Flow, Effect (TypeScript library), TypeScript, XState, Vercel, Netlify, Google Cloud, AWS, Postgres, PlanetScale, Supabase, Cloudflare, Proxmox, Gitea, npm and changesets, Docker, Podman, Discord, Telegram, iMessage, Hermes (agent), Gemini video API, Frame.io (Chuck built a clone), Adobe Premiere, Elgato teleprompter, Google Pixel Watch, Android. ### Projects and sites aihero.dev (skills and courses), Total TypeScript, AI Coding Dictionary, github.com/mattpocock/skills, course-video-manager, Sandcastle, Evalite, Hono (honojs), badass.dev, NetworkChuck Academy. ### Sponsors ThreatLocker Zero Trust World 2027 (ztw.com), Coveron, Bitdefender RealCheck. ### People Matt Pocock, Chuck Keith (NetworkChuck), John Ousterhout, Eric Evans, Michael Feathers, David Thomas, Andrew Hunt, Martin Fowler, Kent Beck, Dex Horthy, Lauren Tan (poteto), Andrej Karpathy (the "LLM wiki" pattern), Elon Musk (as the subject of a deepfake scam stream). ## Biographies **NetworkChuck (Chuck Keith)**: Texas-based IT educator. LinkedIn lists him at NetworkChuck since October 2014 (Dallas), earlier as a CBT Nuggets trainer (2018 to 2020), with a Cisco CCNA from 2013. He told *Raspberry Pi Official Magazine* he sold toilets before studying for CompTIA A+, started in help desk and became a network engineer. His YouTube channel (joined April 2014) had about 5.47 million subscribers on 2026-11-20 per vidIQ. He runs NetworkChuck Academy and a coffee line, and closes videos with a prayer. **Matt Pocock**: British educator and engineer near Oxford. He has a BA in Drama (University of Exeter) and an MA in voice and singing (Guildford School of Acting), and was a voice and public-speaking coach at VoiceHacker from 2015 to 2018. He moved into development in 2018, joined Stately's XState core team in 2021, and did a three-month Vercel developer-advocate contract in 2022 (Turborepo docs) before going full-time on **Total TypeScript**, built with badass.dev. He now runs **AI Hero** (aihero.dev) and the `mattpocock/skills` repo (about 329k GitHub stars on 2026-11-20). His YouTube channel showed about 447K subscribers on the same day. (Sources: AI Engineer speaker bio, LinkedIn, badass.dev, mattpocock.com.) **John Ousterhout**: Stanford computer-science professor (creator of the Tcl language) and author of *A Philosophy of Software Design*, which argues that "the best modules are deep" (quoted in Pocock's README). **Eric Evans**: author of *Domain-Driven Design: Tackling Complexity in the Heart of Software* (2003), which introduced the "ubiquitous language" idea that Pocock's glossary adapts. **Michael Feathers**: author of *Working Effectively with Legacy Code* (Prentice Hall), a classic on getting untested code under test. **Dex Horthy**: of HumanLayer, who popularized the "smart zone / dumb zone" heuristic. In his talks and interviews, quality degrades past roughly 40% of a model's context window, and the fix is frequent intentional compaction and research / plan / implement phases. **Lauren Tan ("poteto")**: engineer at Cursor and SpaceX AI (formerly Meta). Her talk "How I shipped 2,000 PRs last month" and a follow-up Q&A with Pocock describe the "trust ladder," verification skills and agents that merge their own verified PRs. Later reporting gives her personal count as 3,127 PRs in October 2026, on a Grok Bot team shipping about 11,000 PRs a month. **Yusuke Wada**: creator and maintainer of Hono, who closed outside pull requests on 2026-10-05 (see claims check). ## Claims check | Claim (timestamp) | Verdict | Notes | | --- | --- | --- | | Pocock "used to work for" Vercel (1:15:26) | **True** | Developer-advocate contract, Aug to Oct 2022 (LinkedIn; badass.dev). | | Drama degree, voice coach for six years, self-taught around 2015 (1:39:56) | **Mostly true** | BA Drama (Exeter) plus an MA in voice. VoiceHacker coaching ran 2015 to 2018 and badass.dev says he was self-employed six years. His first dev job was in 2018. | | Teacher since 21, "14 years" (0:18:34) | Unverified | Consistent with about a decade of teaching by 2022 (badass.dev). | | Hono now only accepts PRs from collaborators (1:29:13) | **True** | PR #5526 merged 2026-10-05: "Pull requests are limited to maintainers." Wada cited 345 PRs in Jul to Sep 2026 vs 82 a year earlier. Sindre Sorhus did the same on 2026-10-01. | | Skills repo "nearly up to 300,000 stars" (1:29:36) | **Understated** | GitHub showed about 328.9k on 2026-11-20. The interview was recorded before Nov 9. | | YouTube "recently crossed 400,000 subscribers" (2:00:35) | **True** | Channel page showed about 447K on 2026-11-20. | | Lauren Tan ships "like 2,500 PRs a month" (0:45:20) | **True (self-reported)** | Her talk and Q&A cite 2,000 to 2,500 a month, later 3,127 in October 2026. Many are maintenance or "gardening" PRs, not features. | | Smart zone / dumb zone idea comes from Dex Horthy (0:36:21) | **True** | Horthy's talks define the "dumb zone" past about 40% context usage. Pocock's 150k-token threshold is his own rule of thumb. | | Attention cost grows quadratically with tokens (0:26:14) | **True** | Standard self-attention scales quadratically with sequence length. That this drives "dumber decisions" is a heuristic, not a law. | | "Opus has a 1 million token context window"; Opus 5.5 is "rapid" (0:25:44, 0:32:26) | Not checked | Model specs were not verified for this note. | | Tweet limit "is 240 or something" (0:55:41) | **False (hedged)** | X's standard limit is 280 characters (paid accounts allow longer posts). He said he couldn't remember. | | *Working Effectively with Legacy Code* "is from 2003" (1:37:03) | **Close** | Usually dated 2004 (Prentice Hall). | | GitHub Actions is "free for open source repositories" (1:04:21) | **True** | Standard GitHub-hosted runners are free for public repositories. | | Evalite has "fallen by the wayside" (1:27:40) | **True** | Its GitHub repo now says "Deprecated and no longer maintained. Langfuse is the recommended replacement." | | Zero Trust World is Feb 17 to 19, 2027 (0:11:36) | **True** | ThreatLocker lists Feb 17 to 19, 2027 in Orlando (Loews Universal Orlando Resort). | | Bitdefender: survey of 7,000 people, 1 in 7 fell for a deepfake (1:18:59) | Sponsor claim, unverified | | | Coveron: up to $2M identity-theft and $100K fraud coverage, 71% off (1:00:31) | Sponsor claim, unverified | | | "Millions of users" of his skills (1:22:34) | Unverified | Plausible given the star count, but there is no public usage figure. | ## Sources - Video metadata, chapters and description: YouTube via yt-dlp. On-screen pages and cards from sampled frames. - github.com/mattpocock/skills README and SKILL.md files (grill-me, grill-with-docs, improve-codebase-architecture, retro); aihero.dev/skills (install commands, changelog v1.3 dated Oct 5, 2026). - GitHub repo pages: mattpocock/sandcastle, mattpocock/evalite, mattpocock/course-video-manager; star count from github.com/mattpocock/skills (2026-11-20). - Matt Pocock: ai.engineer speaker bio; LinkedIn (mapocock); badass.dev "Launching Total TypeScript"; mattpocock.com; totaltypescript.com. - NetworkChuck: LinkedIn (chuckkeith); vidIQ channel stats (2026-11-20); networkchuck.com/about; *Raspberry Pi Official Magazine* interview. - Hono: Yusuke Wada's HonoConf 2026 and FOSS4G 2026 slides; Reveneau report on PR #5526; honojs/hono PR #5593. - Lauren Tan: TextPurr transcript of her talk; mwolf.dev write-up of the Pocock Q&A; XiaoHu.AI report on AI Engineer CODE 2026. - Dex Horthy: The Pragmatic Engineer ("Context engineering with Dex Horthy"); AI Engineer talk "No Vibes Allowed." - Zero Trust World: ztw.com and ThreatLocker events page.