After pstack, Theo pointed the same audit prompt at Pocock's repo and got a second ranked document (13:07).

The engineering directory

The directory that matters is skills/engineering, containing (13:30):

ask-matt · code-review · codebase-design · diagnosing-bugs · domain-modeling · grill-with-docs · implement · improve-codebase-architecture · prototype · research · resolving-merge-conflicts · setup-matt-pocock-skills · tdd · dissect · tickets · triage · wayfinder · wizard

Productivity skills sit alongside: grill-me, handoff, two-questionnaire, teach, wait-what, writing-for-agents (23:30). Reference skills: codebase-design (vocabulary for designing deep modules) and domain-modeling (interpreting a project's own words and writing them down) (23:30).

User-invoked vs model-invoked

The structural detail Theo rates highest in the collection (21:03):

  • Many of Pocock's skills set disable model invocation, meaning the model can't pull them in on its own — you invoke them manually with a slash or $ command. Theo calls this "a very good call" given how some of them behave (14:08).
  • The docs separate user-invoked from model-invoked, and Theo emphasizes these are genuinely different things:
    • User-invoked — effectively plugins you call when you want them (grilling, teach, wait-what, wizard).
    • Model-invoked — steering the model to work differently on its own (diagnosing-bugs).
  • Example of correct classification: diagnosing-bugs is deliberately not manual-only. It fires by itself during debugging, and Theo confirms it has been firing and helping over several days — his agents find root causes and communicate what's wrong more effectively (14:37, 14:57).
flowchart TD
    S[Skill in the repo] --> Q{disable model invocation?}
    Q -->|Yes| U[User-invoked<br/>slash / $ command<br/>grilling · teach · wait-what · wizard]
    Q -->|No| M[Model-invoked<br/>fires on its own<br/>diagnosing-bugs]
    U --> UP[Acts like a plugin you reach for]
    M --> MP[Steers how the model works]

Skills that call other skills

  • Pocock's skills are self-referential: improve-codebase-architecture tells the agent to run codebase-design first to acquire the vocabulary, then presents candidates as an HTML report — a touch Theo liked (14:08, 20:38).
  • It also pulls in a CONTEXT.md vocabulary for the domain and calls out grilling and domain-modeling (20:38).
  • grill-with-docs is itself thin — it's a relentless interview that produces docs, ADRs and a glossary as it goes, mostly by composing other skills (14:08).

The prescribed main flow

Theo eventually finds the getting-started docs ("apparently I should have been reading the docs the whole time"). Entry point is a setup-skills command that configures issue tracking so the skills can drive it (22:20).

flowchart LR
    A[grill-with-docs<br/>interview, record decisions] --> B[spec]
    B --> C[tickets<br/>Jira / Linear / GitHub]
    C --> D[implement]
    D --> E[code-review]

Supporting entry points around that spine (22:38, 23:03):

  • shaping — turn vague ideas into executable plans.
  • wayfinder — chart a large effort as a map of decisions, then settle them.
  • prototype — mock designs to find out what you want before shipping; explicitly deletable code.
  • research — cited answers from primary sources. Theo is the least sold on this one: he thinks agents already verify claims well enough if you just ask (23:03).

The recurring complaint: em dashes

  • Theo counts nine em dashes on the grilling skill page alone — at least one per paragraph in most paragraphs (15:23).
  • He connects it causally to his own output: "I was wondering why I was getting em dashes again. It might be Matt's skills." (15:23)
  • Half-joking, he floats forking the entire repo with only the em dashes stripped out (30:06).