← narwal.one/Second Brain
SecondBrain
Ask the Brain
Index/Sourceupdated Sat Sep 26 2026 08:00:00 GMT+0800 (Philippine Standard Time)

AI Skills with Matt Pocock (The Pragmatic Engineer)

matt-pocockskillsgrill-mewayfindercontext-windowstrategic-programmingsoftware-fundamentalstddagentsai-engineeringeducationpragmatic-engineer

AI Skills with Matt Pocock (The Pragmatic Engineer)

~96-minute long-form interview: Gergely Orosz (The Pragmatic Engineer) with Matt Pocock. Half career story (voice coach → self-taught dev → XState core team → 3-days-a-week Vercel contract → Total TypeScript), half the most complete account yet of how Pocock's skills fit together as a workflow — and why he keeps finding the answers in 20–50-year-old software books.

Where his two AI Engineer talks (Building Great Agent Skills (Matt Pocock, AI Engineer), Software Fundamentals Matter More Than Ever (Matt Pocock, AI Engineer)) gave the rubric and the thesis, this conversation gives the operating model: grill → spec → tickets → AFK implement loop, sized to the model's smart zone, with a human doing the strategic work.

Key claims

  1. "AI has largely eaten tactical programming; it's up to us to handle the strategic." Borrowing Ousterhout's split (Strategic vs Tactical Programming). Knowledge (syntax, the what) is now nearly free; wisdom (the why) has gotten no easier to learn — and is where his educational business survived (Total TypeScript revenue is down; the strategic-layer AI course is working).
  2. The workflow is a context-window budget. Following Dex Horthy's Smart Zone and Dumb Zone (~first 150k tokens of any window size are the smart zone), all large work is split across sessions: Grill Me (align) → spec (the "destination document"; he used to call it a PRD) → tickets (one per session) → an implement loop run AFK. Ralph loops were an earlier version of the same idea — smallest change toward the goal, clear context, repeat; state lives in the file system, not the model.
  3. Wayfinder for work too big to grill in one session. A map + fog of war + tickets of different types (grill, prototype, research, arbitrary task); maps reach 50–100 tickets. He uses it for non-software work too — course planning, building a garden office.
  4. Decision rule for how much upfront alignment — "shift right as much as possible":
    • Small enough to align after seeing it (bug fix, 3-pixel nudge)? → no grilling; feedback button → GitHub issue → implement agent → review agent → he aligns on the result.
    • Fits in one session but hard to row back from? → Grill Me.
    • Spans multiple sessions? → Wayfinder. Orosz notes this reproduces big-tech's pre-AI PRD sizing heuristic (one-day project: just build; one-month project: 1–2 days of planning).
  5. Day shift / night shift. Plan with the human during the day; agents implement during the night; review clean code in the morning. Motivation: escape the "100 terminals, boom-boom context switching" pattern in favour of long focused chunks.
  6. Leading Words come from old books. Tracer bullets and vertical slices (Pragmatic Programmer), deep modules (Ousterhout), ubiquitous language (Evans DDD) are all in the model's prior — say them and the agent repeats them in its reasoning traces. Orosz's parallel: jargon is compression between professionals; Kent Beck and Ward Cunningham kept a thesaurus on the desk to find the right word.
  7. "Momento-driven development." An agent is a new starter every morning with no memory; humans work around a bad codebase by accumulating memory, agents can't. So optimise the codebase for new starters — which is what software fundamentals always asked for. The codebase is the environment the agent operates in.
  8. Why agents excel at software: inputs and outputs are all text (code, docs, test/type/lint results). Anything non-text (UI interaction feel, animation) is still "garbage". Generalisation: make your work text-based / agent-friendly and agents will do well in any discipline.
  9. Selling fundamentals to non-engineering stakeholders: (a) observability over every agent — success/failure rates per repo, which would be invasive for humans but is fine for a paid service; (b) someone owns that data; (c) a common org-wide skill set so teams can contribute back and A/B test workflows; (d) consider ~20% of time on "the factory that builds your software". See Software Factory.
  10. Local → cloud agents. "Moving away from my local dev setup makes zero sense to me." A remote always-on box he chats with in Discord (schedules a morning stand-up, works while he's on the train); collaboration (tagging a colleague into a grilling session) needs shared spaces (Slack/Discord/Teams/Linear), not 100 private terminals. Orosz: Ramp/Stripe/Uber platform teams see 70–80% of devs voluntarily choosing cloud dev environments (frontend is the exception).
  11. TDD, reconsidered. TDD optimises for small human working memory; agents have large working memory, so TDD "is sort of aiming at the wrong problem." What agents do need is feedback loops and proof — "give me TDD evidence that it would fail without this change" — which is also hard for the agent to cheat. Agents also write tautological tests. See contradiction below.
  12. Tech debt in tiny codebases. Reply to YC's Jared Friedman ("tech debt used to be something you lived with in a large codebase"): "now you can live with it even in a tiny code base." Mitigation: implement agent + automated review agent enforcing standards — but then who reviews the reviewer? A permanent battle requiring strategic judgment.
  13. "The only thing your team needs are gardeners." We are the agents' platform team; the essential skill is diagnosing entropy before it bites. The trait of great engineers in one word: introspection — putting your own process into words the AI can execute (example: Lars Grammel building a software factory for the Vercel AI SDK).
  14. Juniors: Pocock would "love to be a junior right now" — use agents as much as possible and stay interested in the process. But he also asks the hard question: if strategic knowledge is where the leverage is and tactical work "has gone below minimum wage", why would a company hire someone without it? Uncle Bob's answer (treat a junior like an agent for a while) he calls "an enormous waste of money". Unresolved — see Deskilling Trap (Juniors).
  15. Education still needs humans for curation: information is a dependency graph; the teacher's job is finding the best linear path through it (his "Dijkstra's algorithm" metaphor) — strategic work AI isn't good at.

Contradictions and updates vs prior Pocock sources

warning Contradicts Software Fundamentals Matter More Than Ever (Matt Pocock, AI Engineer): that talk put TDD at the centre of failure-mode #3 ("rate of feedback is your speed limit… TDD to force small steps"). Here Pocock says TDD "optimises for a very small working memory… agents don't need that… it's aiming at the wrong problem", and he now prefers asking for proof/TDD evidence over strict red-green-refactor. He still recommends his TDD skill for human confidence. Read as a refinement: feedback loops remain the constant; TDD is demoted from method to one way of producing evidence.

warning Contradicts Grill Me / Matt Pocock (vault lineage): the vault framed Grill Me as the ancestor of the interview-me pattern that Thariq Shihipar later recommended. Pocock says the opposite: "I think I got it originally from like a Tariq who works with Claude Code… get the agent to interview you." Shihipar → Pocock, not Pocock → Shihipar.

note Venue detail: Pocock places his "Software Fundamentals Still Matter" talk at AI Engineer London (April), now ~1.2M views; the vault's source page labels it AI Engineer World's Fair 2026. Unverified which recording the vault's page captured — possibly the same talk delivered twice.

  • Scale update: the mattpocock/skills repo is now ~230,000 stars — "second most starred skills repo in the world", ~20th–25th most starred repo of all time (vault previously recorded ~13,000 stars for Grill Me alone).
  • SDD nuance: he has "mixed feelings" about spec-driven development ("a strange term, it encompasses too much") and repeats the specs-to-code failure story — yet his own pipeline centres on a spec as destination document. Consistent with the vault's reading on Spec-Driven Development (SDD): he rejects spec-as-source-code, not specs.

Enterprise / leadership reading

  • The workflow is an org design pattern, not just a personal hack: humans own strategic alignment (grill/spec/map), agents own tactical execution in bounded sessions, review agents police entropy, and a platform/"gardener" function owns the factory. That maps onto the vault's Software Factory and Agentic Pods threads.
  • The stakeholder argument for fundamentals is now measurable in a way it never was for humans: agent observability → per-repo success rates → evidence that codebase health drives agent yield.
  • The junior-hiring question is posed more sharply than anywhere else in the vault: strategic knowledge is the lever, and the traditional way of acquiring it (years of tactical work and battle scars) is being priced out.

Entities & concepts

Source