メニュー
Self-check Markdown changes under docs/ against the reposix progressive-disclosure layer rules before committing. Use whenever you author or edit a file in docs/index.md, docs/concepts/, docs/tutorials/, docs/guides/, or docs/how-it-works/.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Use as a reposix orchestrator/coordinator facing one of five concrete situations, unsure whether to act, wait, or escalate — (1) a dispatched child hasn't committed in a while, stuck vs slow?; (2) about to dispatch a fix for a BLOCKER/HIGH whose only evidence is a code reading, not something executed; (3) an intake/ledger row has a fix design already sketched, about to implement as written; (4) the next roadmap item no longer seems right given new information; (5) found out-of-charter work (bug, missing tool, doc lie) and unsure whether to fix now, file, or ask. Also for the fable-consult template or the CONSULT-DECISIONS.md ledger format. NOT for routine dispatch (use coordinator-dispatch) or anything a hook already resolves.
Dispatch template for a reposix phase/debt-drain coordinator — charter to paste, canonical role→subagent_type mapping, model-tier table, lane-slicing checklist, report contract, relief trigger, pause/resume brief. Pull this before dispatching a wave of sub-agents.
Run subjective rubric checks (cold-reader hero clarity, install positioning, headline-numbers sanity) by dispatching one unbiased subagent per stale catalog row in parallel. Reads quality/catalogs/subjective-rubrics.json, persists JSON artifacts to quality/reports/verifications/subjective/<rubric-id>.json. The next runner sweep re-grades the catalog rows. Pre-approved for v0.12.0 P61 per .planning/research/v0.12.0/open-questions-and-deferrals.md line 124.
Spawn the reposix dark-factory regression test — a fresh subprocess agent given only `reposix init` and a goal must complete clone+grep+edit+commit+push using pure git/POSIX, including the conflict-rebase and blob-limit recovery cycles. The StrongDM/dark-factory regression harness for the v0.9.0 architecture. Invoked from CI release-gate and from local dev (/reposix-agent-flow).
Run the docs-alignment dimension — refresh stale rows for a single doc OR run the full backfill audit across docs/ + archived REQUIREMENTS.md. Subagents propose claim→test bindings with file:line citations; the `reposix-quality` binary validates and mints catalog state. Reads quality/catalogs/doc-alignment.json. Top-level only (depth-2 subagent fan-out unreachable from inside /gsd-execute-phase; subscription users cannot fall back to claude -p).
Run the full docs-alignment backfill audit across docs/ + archived REQUIREMENTS.md. Top-level slash command — delegates to reposix-quality-doc-alignment backfill playbook. Dispatches ~25–35 Haiku extractor subagents in waves of 8, runs merge-shards, generates PUNCH-LIST.md.
| name | reposix-banned-words |
| description | Self-check Markdown changes under docs/ against the reposix progressive-disclosure layer rules before committing. Use whenever you author or edit a file in docs/index.md, docs/concepts/, docs/tutorials/, docs/guides/, or docs/how-it-works/. |
This skill encodes the P1 + P2 framing rules from
.planning/archive/notes/phase-30-narrative-vignettes.md so an authoring agent can
self-check before commit. The mechanical enforcer is
scripts/banned-words-lint.sh; the canonical config is
docs/.banned-words.toml.
Before staging any change under docs/, run:
scripts/banned-words-lint.sh # default — Layer 1 + Layer 2
scripts/banned-words-lint.sh --all # QA — also Layer 3
Pre-commit and CI both run the default mode automatically. Running it locally
catches violations during authoring instead of after git commit.
| Layer | Pages | Banned vocabulary |
|---|---|---|
| 1 — Hero | docs/index.md | the v0.1 FUSE-era plumbing words (FUSE, fusermount, inode, daemon, kernel, syscall) and the v0.9.0 git-native plumbing words (partial-clone, promisor, stateless-connect, fast-import, protocol-v2) |
| 2 — Below the fold | docs/concepts/, docs/tutorials/, docs/guides/ | same as Layer 1 |
| 3 — How it works | docs/how-it-works/ | nothing layer-specific (P1 still applies) — plumbing terms are permitted; this is where the technical reveal happens |
| 4 — Reference + Research | docs/reference/, docs/research/, docs/decisions/ | nothing |
P1 (banned everywhere): replace — reposix complements REST; it does not replace it.
If a banned word is genuinely required (a quote, an error message, a
historical note), append the marker <!-- banned-words: ok --> to the same
line. Pair the marker with a brief comment explaining why; reviewers will
push back on bare allowlists.
Example:
The deleted v0.1 design used a FUSE daemon. <!-- banned-words: ok --> historical
docs/.banned-words.toml.scripts/banned-words-lint.sh
(the script flags this duplication on purpose — drift between the two is a
review surface).docs/.banned-words.toml, the script, and the doc fixes together.Above the fold, the only technical words permitted are ones every developer already knows — file, folder, edit, commit, push, merge, YAML, markdown. —
.planning/archive/notes/phase-30-narrative-vignettes.md§"P2 progressive disclosure"
The linter is the institutional memory of that rule, not a checklist (per project CLAUDE.md OP-4: "ad-hoc bash is a missing-tool signal — promote it to a committed artifact").