Skill: glossary

Lazy utility that fixes the meaning of a domain term in CONTEXT.md the moment it first surfaces, so its sense doesn't drift across the pipeline. For each term it captures a one-sentence canonical definition and — when the word is ambiguous — a NOT-reference naming the concept it's confused with. Runs anytime, with no upstream gate: a single term mid-interview, an undefined-term finding clarify resolves mid-sweep, or a batch handed over by specify. The output feeds specify (role + domain-term names) and design (invariants), which treat ## Glossary as canonical and override anything that contradicts it.

This is a capture utility, not a Socratic stage — it does not run the shared Socratic loop or critic. The one shared dependency is question phrasing: → ../_shared/ask-style.md

Term definitions follow artifact_language — the ## Glossary heading and the other H2s stay English, the frontmatter stays the template's (status: Living + updated_at: — never write artifact_language or any other settings key into the file) → ../_shared/artifact-language.md.

Owner

Whoever drives the conversation — anyone who spots ambiguity. Tech Lead approves the canonical form when a term is contested.

Inputs

• <term> — the domain word/phrase to fix. If not given, ask for it.
• (Optional) <slug> — feature slug; targets docs/features/<slug>/CONTEXT.md. Absent → repo-root CONTEXT.md.
• (Optional) pending_glossary_terms — a batch handed over by specify after it writes the spec. Process each term in turn.
• (Optional) an undefined-term finding from clarify — clarify invokes this skill mid-stream, one term at a time, the moment such a finding is Resolved (it never invents a meaning inline; the canonical definition lands here).
• (Optional) proposed phrasings already surfaced in an interview/brainstorm — offer them as definition options instead of asking blank.

Protocol

1. Pick the target (the two-level contract). Root CONTEXT.md = project-wide terms (meaningful across features); docs/features/<slug>/CONTEXT.md = feature-scoped terms (meaningful only inside that feature). If <slug> is given → the per-feature file; otherwise → repo-root. A term lives in exactly one of the two files — never split or duplicate it across levels. Readers (specify, clarify, design, api) read both files; on a conflict the per-feature entry wins.
2. Generic-term filter. Reject words that name infrastructure or transport rather than the business domain (e.g. HTTP, queue, cache, the datastore, the broker, a framework). Refuse: «<term> is technical, not a domain word — its choice belongs in the SAD or an ADR, not the glossary». Continue only for genuine domain vocabulary.
3. Bootstrap (lazy). test -f <target>. Missing → copy ./templates/CONTEXT.md to <target>. Present → read it.
4. Conflict check — both levels. grep -i "^- <term>" <target> AND the other level's file when it exists (root CONTEXT.md ↔ docs/features/<slug>/CONTEXT.md).
   • Found (either level), identical sense → STOP, report «already in the glossary» + which file holds it — never duplicate a term across levels.
   • Found (either level), different sense → escalate via AskUserQuestion (phrasing per ../_shared/ask-style.md): «<term> is already defined as <existing> in <file> — same concept or a different one?». If different: a genuinely feature-scoped narrowing goes to the per-feature file (readers let it win there); otherwise propose a disambiguating pair of names (e.g. a billing-scoped vs a runtime-scoped variant) and fix both.
   • Not found in either → continue.
5. Ask the canonical definition — one AskUserQuestion: «Define <term> in this domain, in one sentence». Offer interview/brainstorm phrasings as options when available; otherwise free text.
6. Ask the NOT-reference — one AskUserQuestion: «Which concept does <term> get confused with, so a future reader doesn't mix them up?». No plausible homonym → None.
7. Compose one line. - <term> — <one-sentence definition>. NOT <confused concept + how it differs>. — or, when step 6 = None, - <term> — <definition>.
8. Append under ## Glossary. Read the file, insert the line in ## Glossary (alphabetical if the section is already sorted, else at the end). Never rewrite existing entries.
9. Prune empty H2s. On a fresh bootstrap, delete ## Invariants / ## Out of scope if they hold no real content — only ## Glossary is mandatory. (A genuine invariant or out-of-scope note goes in its section, never as implementation detail.)
10. Structural self-check — per ../_shared/self-check.md: re-read the target file(s) from disk and verify 3 items: (1) the term appears exactly once across both levels combined (grep root CONTEXT.md + docs/features/<slug>/CONTEXT.md — no duplicate, no split); (2) the entry matches the format - <term> — <definition>. (with the optional NOT … tail); (3) no empty H2 sections remain. Fix + re-check ≤2 cycles; surface anything unresolved.
11. Stamp + commit + handoff. Set updated_at: <today> in the frontmatter. Propose context: + <term> (or context: + <term>, <term2> for a batch). May fold into the caller's intake/spec commit. Then emit the stage-handoff block per ../_shared/handoff.md (utility variant) — What I did (incl. «self-check: 3/3 pass») + Review (CONTEXT.md) + Run next: resume your backbone stage (e.g. /sdd:design <slug>); /clear optional.

Definition of Done

• <target> exists and contains <term> under ## Glossary in the «one-sentence canonical + optional NOT-reference» format.
• Any conflict with an existing entry is resolved (reported as duplicate, or disambiguated into distinct names).
• Generic tech words are refused, not stored.
• Empty H2 sections are pruned on bootstrap; ## Glossary remains.
• updated_at reflects today; a commit is proposed.

Anti-patterns

• Glossary as spec or scratch pad. Implementation detail («counter stored with a 1-minute TTL») belongs in the SAD/ADR, not here.
• Empty H2 «for completeness». A heading with no bullets — prune it.
• Silent edits. Adding a term without confirming the definition; the author must control the glossary.
• Batched «I'll add them later». Capture each term the moment it surfaces — deferral loses it.
• Storing generic tech words. HTTP, queue, the datastore name — refuse them.
• Rewriting on re-run. The skill reads and appends; it never overwrites the file.
• Ambiguous term with no NOT-reference. A homonym without «NOT …» is a six-month confusion waiting to happen.

References & template

• ./templates/CONTEXT.md — output scaffold; inline comments are the per-section contract (Glossary mandatory, Invariants/Out-of-scope pruned when empty).
• ../_shared/ask-style.md — phrasing for the definition / NOT-reference / conflict questions.

Example invocation

User: «add term tenant for rate-limiting-per-user» Skill: <slug> given → target docs/features/rate-limiting-per-user/CONTEXT.md. Generic filter: tenant is domain → continue. File missing → copy template. grep "^- tenant" → not found. Definition Q → «a billable customer organisation owning 1+ users». NOT-reference Q → «NOT user — a user is one person inside a tenant». Compose - tenant — a billable customer organisation owning 1+ users. NOT user (a user is one person inside a tenant). → append under ## Glossary → prune empty ## Invariants/## Out of scope → updated_at: 2026-05-28 → commit context: + tenant.