Grill skills for design interviews + ADR-0009 (no self-model)

[jbeda]

Two independent changes that came out of the same session: agent skills for design interviews, and the ADR that one of those skills leans on.

chore(skills) — grilling and grill-with-docs

Adds .claude/skills/, for skills used when working on modelith (distinct from plugin/skills/, which ships to users of the CLI).

• grilling — the interview engine, taken verbatim from mattpocock/skills (MIT). One question at a time, look facts up rather than asking, put decisions to the user, don't act until shared understanding is confirmed.
• grill-with-docs — same interview, but records what comes out of it. Rewritten for this repo rather than copied.

Upstream's version of the second skill maintains a CONTEXT.md glossary and carries its own ADR-format doc. Both are dropped here, because this repo already has stronger versions of each: the format's vocabulary lives in the schema, the structs, and the schema reference (two of which TestSchemaStructSync holds in sync), and .claude/rules/adr.md already sets the ADR bar. What the skill keeps is the capture map:

Grilling produces	Lands in	During the session?
Hard-to-reverse decision, real trade-off	project-docs/adr/ per .claude/rules/adr.md	Write it inline
Format or vocabulary change	schema + structs + schema reference, as one change	Flag only — they're test-coupled; a half-written edit leaves task check red with the design still being argued
Running state	HANDOFF.md	End of session

Provenance. .claude/skills/SOURCES.md records the upstream source, the commit it was taken at, the licence, and what changed locally — and sits outside any skill directory so it never loads into agent context. CLAUDE.md gets a line making that the convention for future imports.

docs(adr) — ADR-0009: modelith is not modelled in modelith

Modelling modelith's own domain in a *.modelith.yaml is the obvious dogfooding move, and the question keeps resurfacing — which is the signal to write it down rather than re-argue it.

The decision is not to. The vocabulary a self-model would capture is already defined in three places, two of them machine-checked against each other. A self-model would join them as a fourth copy that nothing validates, and drift between the tool's own model and its own schema is the worst possible advertisement for the tool. It would also compete with examples/ for the worked-example job, and invite edits made for demo reasons to a file that is a golden fixture.

Pinned by TestADR_0009_NoSelfModel: the set of committed *.modelith.yaml files must be exactly the two known examples, so a new one fails the test until someone adds it to the allowlist — which puts the ADR in front of them. Verified to fail on an added root-level model, not merely to pass on the status quo.

Notes for review

• task check is green.
• No /code-review loop was run. Per .claude/rules/agent-workflow.md the loop is available discipline for changes that put a correctness-critical surface at risk; this is prose plus one guard test, and the test was verified in both directions by hand.
• The two commits are independent — the skills change and the ADR can be split or dropped separately if you'd rather.

🤖 Generated with Claude Code