define-domain

Define Domain

Actively build and sharpen the project's glossary while designing. Merely reading CONTEXT.md for vocabulary is a habit any skill has; this skill is for changing the model — challenging terms, probing edges, and writing settlements down the moment they happen.

During the conversation

When another skill owns the interview (e.g. clarify-decisions inside frame-change), you run as a passive side effect, not a second interviewer. Do the silent work directly — record a settled term to CONTEXT.md, cross-reference the code — but anything that needs the user (a challenge, a disambiguation, an edge-case probe) does not get its own turn: hand it to the interview's single question channel as the next question, asked one at a time in dependency order. Only when you are invoked standalone do you drive the questions yourself.

Challenge terms against the glossary. When the user's usage conflicts with CONTEXT.md, call it out on the spot: "Your glossary defines cancellation as voiding the whole order, but you seem to mean removing one line item — which is it?"

Sharpen fuzzy language. When a term is vague or overloaded, propose a precise canonical one: "You keep saying account — is that the Customer or the User? They differ here."

Stress-test with edge-case scenarios. When relationships between concepts come up, invent concrete scenarios that press on the boundaries: "A Member leaves mid-billing-cycle with an unpaid Invoice — what happens to their Workspace?" Force the boundary to be stated, not implied.

Cross-reference the code. When the user asserts how something works, check whether the code agrees. Surface contradictions explicitly: "The glossary (and you) say partial refunds exist, but the code only refunds full payments — which is right?" Never let the glossary and the code silently disagree.

Updating CONTEXT.md

Update CONTEXT.md inline, the instant a term is settled — never batch glossary edits for later; batched terms get forgotten. If no CONTEXT.md exists yet, create it lazily when the first term settles (seed from the skill set's templates/CONTEXT.md. Resolve pack seeds in this order, first path that exists: (1) templates/ beside this SKILL.md, (2) ${CLAUDE_PLUGIN_ROOT}/templates when that variable is set, (3) ../../../templates relative to this SKILL.md).

CONTEXT.md holds exactly one kind of entry: a term definition. Decisions go to ADRs; implementation detail and spec fragments go to the specs. Term format:

**Workspace**:
The top-level container a team collaborates in; owns all Notes and Folders.
_Avoid_: project, space, tenant

Be opinionated: one canonical term, competitors banished to the _Avoid_ list. Definitions stay to one to three tight lines — what the thing is, not how it's built. Only project-specific concepts belong; general programming vocabulary does not.

ADRs — sparingly, through the three-part gate

Offer an ADR only when all three hold:

1. Hard to reverse — changing course later carries real cost.
2. Surprising without context — a future reader would ask "why on earth this way?"
3. A real trade-off — genuine alternatives existed and one was chosen for specific reasons.

Any one missing → no ADR. For contrast — a decision that is surprising and a real trade-off but cheap to reverse earns none: a 200 ms debounce on the search box trades responsiveness against request volume, yet flipping it later is a one-line change, so it fails "hard to reverse." Record it in the code, not an ADR. Format: docs/adr/NNNN-slug.md (create the directory lazily), numbered sequentially — scan for the highest existing number and add one. Body is a short title plus 1–3 sentences: context, decision, why. That's the whole document; recording that and why is the value, not filling sections.

If a new decision contradicts an existing ADR, flag the conflict to the user and resolve it explicitly (supersede the old ADR by number).

When pruning or auditing ADRs

The three-part gate decides whether to write an ADR. This section decides whether a live one still earns its place.

Word count, age, and folder size are not archive criteria. A lead, a review tomorrow, or a quota does not replace classification.

Classify every live file. Done when: each one is keep, archive, or drop, with a one-line reason.

• Keep — a future change still needs it: the decision still constrains the running system, or a rejected alternative is still being pitched and this file is why it lost.
• Archive — the decision shipped and the body will not guide a future change (closed UI chrome, implementation recap, process history now obvious from the code). Move it to docs/adr/archived/ and insert Archived: YYYY-MM-DD under the title. Do not rewrite the body.
• Drop a rejection only when the losing idea is no longer a plausible proposal. Otherwise keep it.

Tightening the folder by hiding a live constraint or a still-tempting rejection is not cleaning. Done when: the live tree still holds every decision that binds or still guards a live bad idea.

Red Flags

• Archiving every file older than N days or longer than N words
• Emptying docs/adr/ so a review "looks tight"
• Treating "don't get precious / the point is the count" as a reason to skip classification

Thought	Reality
"Judging each file against still-useful is the opposite of the lead's explicit criteria"	The quota is a discovery hint, not the test. Classification is the job
"A tight live folder by mechanical rule is what was ordered"	A tight folder that hid the decisions still in force is a hollow review
"Don't get precious about individual files; the point is the count"	The point is which decisions still bind. Count is a side effect