conversion-inventory.md

Conversion inventory — every corpus rule, audited

Companion to ../engine/checks/DESIGN.md (issue #127). Each discrete instruction in the mounted corpus, classified by the mechanism that should carry it. This is the audit that answers "is enough of the corpus convertible?" — and the worklist Phases 1–3 execute against.

Dispositions:

• check — post-hoc deterministic check (pack named). (adv) = advisory by kind — the rule's semantics are directional (a smell to judge, not a defect to fix); everything else ships blocking, fail-fast.
• hook — PreToolUse gate, blocks before the action runs.
• skill — becomes an on-demand procedure; unloads from always-on context.
• setting — platform enforces it (GitHub repo/branch settings); no code at all.
• stays — process/judgment guidance; remains an instruction. (judge) = a deferred LLM-judge candidate. A whole doc (or a doc's residue after its checks are carved out) that stays is delivered as a skill per DESIGN.md's catalog — same content, harness-managed trigger instead of an index line the agent must remember to follow.
• knowledge — platform gotcha; content unchanged, no enforcement possible — delivered as a paths-/description-triggered skill.

Verdict up front

~125 discrete instructions audited. ~45 convert to deterministic mechanisms (≈40 checks across five packs, 4 hooks, 2 settings, plus two whole docs becoming command skills); ~60 stay as instructions (process/judgment) and ~20 are knowledge — of which roughly half carry a static signature in the artifact and convert into tech-pack checks (Phase 2), the failing check teaching the gotcha at write-time instead of runtime. What genuinely stays no longer means an index soft pointer either: the stays docs and signature-less residues are delivered as skills (DESIGN.md's catalog), trading the remembered trigger for a harness-managed one at roughly the same token cost. So: a majority of instructions enforced mechanically — no, ~40%. A majority of the always-loaded context burden eliminated — yes: four of the five force-loaded docs unload entirely, the two most breakage-prone task docs (filePlacement, textAndFileManipulation) shrink to rationale, and the routing index largely dissolves into skill descriptions. Force-loaded prose that remains: trimmed working-discipline and the engineeringPractices judgment core.

always/ — the force-loaded baseline (highest standing cost)

Rule	Disposition
working-discipline: problem-first consensus before implementing	stays
working-discipline: confirm behavior isn't already provided	stays
working-discipline: misread ≠ wrong artifact; size writing to its idea	stays
working-discipline: clean-room rebuild from the authoritative source	stays
working-discipline: fix warnings, don't tolerate	per-project CI config (fail on warnings); rule text leaves baseline
working-discipline: never suppress a warning as the quick path	check basics/warning-suppression: a eslint-disable/@ts-ignore/noqa-style marker with no reason at the site (inline or in the comment immediately above) → finding; a self-documented suppression passes
working-discipline: approval applies only backward	stays
task-lifecycle: issue before work; commits reference it	check basics/task-lifecycle: commits since merge-base carry #N; the issue exists
task-lifecycle: update issue status as work progresses	check (adv), partial — full flow lives in the merge skill
merge-to-main: the recipe (~4 calls)	skill (trigger stays the owner's "LGTM" preference)
merge-to-main: squash as the method	setting — GitHub "allow squash merging" only — verified by check basics/squash-merge-history: the change introduces no merge commits (scoped to the work — its own commits since the merge-base — not the repo's whole history), plus a CI-surface config check later
merge-to-main: gate on CI only if the repo has it	folds into the skill
merge-to-main: lessons pass on every merge	folds into the skill (deterministic step, not a remembered trigger); later became the conversation capture step — extraction moved to the growth pack's nightly
merge-to-main: don't re-read the issue; don't fight branch deletion	folds into the skill
temporary-workarounds: never delete a remote branch	hook — PreToolUse blocks git push delete refspecs; doc deleted when its list empties
extracting-lessons.md (entire method doc, force-loaded; now packs/grow_with_claudinite/extracting-lessons.md)	skill at conversion, later a growth-pack method doc — applied by the conversation-extract daily task (and any owner-requested retrospective); never always-loaded again

Baseline outcome: five force-loaded docs → working-discipline (trimmed) plus a slimmer index.

skills/file-placement/SKILL.md — the flagship conversion

Rule	Disposition
Reference-distance metric (0/1 healthy, 2+ reach; edge-counting)	check basics/file-placement (adv) — the metric is already specified mechanically
Mandated-location exemption (.github/, .claude/, root manifests)	codes into the check as an exemption list
Keep the mandated file thin (launcher → one near entry point)	check (adv): count of distinct deep references from a mandated file
Test-location-convention exemption	codes into the check: detect the project's test root, exempt test→tested
Mirror the source tree when picking a test convention	check (adv): source file with no mirrored test path
Deliberate cross-cutting concerns are allowed, few and named	the acceptance mechanism, with reason strings
Tooling acts on paths → structural, fail-safe splits	stays (judge)
Doc's remaining role	shrinks to rationale + "what to do about a smell"

skills/repo-text-sweeps/SKILL.md — outcomes convert, procedure follows them out

Rule	Disposition
Grep inbound references after delete/rename	check basics/reference-integrity: dangling relative links, imports, index entries
Double-prefix / URL corruption in a naive sweep	outcome caught by reference-integrity + project tests; procedure text shrinks
path.join-segment references survive a rename pass	stays (in-flight; no "old path" known post-hoc) — outcome partially caught by tests
Specific→catch-all half-converted paths	same — outcome caught by reference-integrity where the result dangles
Re-read a file at its new path after git mv	stays (in-flight tool behavior)
Markdown link carries its path twice (label vs. href)	check: label path ≠ target path → finding; fully deterministic
Exclude by pathspec, not grep -v on content	stays (procedure); the runner's own lib obeys it
Derive structural-check file sets from git ls-files	stays; the runner obeys it
Write tool, never cp from harness-internal paths	stays (in-flight)

skills/git-github-advanced/SKILL.md — a lint pack plus two hooks

Rule	Disposition
Status updates: comment, don't issue_write-overwrite	hook — PreToolUse warns/blocks issue_write update with a body
Commit often, in layers; tests-first ordering	stays
Don't rewrite published/shared history	hook (block force-push to protected branches) + setting (branch protection) — the setting verified by a fleet-surface config check (branch-protection reads need an owner token)
Squash-merge branch-sync gotchas (rebase --onto, stale lease)	knowledge
Commit-nag ≠ authorization to commit drift	stays
Open a PR early when the artifact is CI-only	stays
Sync early to keep conflicts small	stays
Fetch before branching off origin/main after a remote merge	knowledge
"Ahead by N" ≠ unmerged (content triage steps)	knowledge
GITHUB_TOKEN push triggers no workflows	knowledge
Gate optional CI jobs on a variable, not a secret	check github-actions/: secrets.* in a job-level if:
Automated job needs a unique branch per run	check github-actions/ (adv): date-keyed branch names in workflows
Create a brand-new label before --add-label	check github-actions/ (adv)
run: steps have no pipefail by default	check github-actions/: piped run: without a bash shell default
Unattended workflow must escalate its own failure	check github-actions/ (adv): scheduled workflows lacking a failure-reporting job
workflow_call permission forwarding	knowledge (lint too fuzzy)
CI reading submodule files needs submodules: true	check github-actions/: repo has .gitmodules → checkout steps need the flag
Mark large fixtures linguist-vendored	check: committed files > threshold in fixture dirs lacking a .gitattributes entry
Renaming a directory housing a submodule	knowledge
Markdown in <td> needs blank lines	check (adv), Markdown lint
commit.gpgsign doesn't cover merges	knowledge
Repo-allowlist scoping: query per-repo, never org-wide	stays (in-flight)
Merging gotchas (relocation refs, invariant reintroduction, porting across invariants)	stays/knowledge — reference-integrity re-run after a merge covers part

skills/engineering-practices/SKILL.md

Rule	Disposition
Name by scope, not technology	stays (judge)
Single source of truth + drift guards	stays — the guards themselves are already tests; the rule to add them is judgment
Shared-constants occurrence guard	already a test pattern in consuming repos; portable pack helper later
A states what from B, never how	stays (judge)
GENERATED filename discipline; never hand-edit or hand-merge	check (adv): diff edits a *GENERATED* file with no generator-input change; .gitattributes merge=ours present
Verify platform behavior against a real run	stays
Earn each dependency	check (adv): diff adds a manifest dependency → finding asks for justification
Bespoke how-to earns its place over upstream docs	stays (judge)
Resilience that swallows errors destroys observability	check (adv): empty/swallowing catch lint
Hook AI-context automation to human workflow events	stays
Expected outcomes are clean exits, not failures	stays
Datacenter-IP-blocked fetches → residential proxy	knowledge
Setup script must cd into the checkout	knowledge
Fresh-checkout Cannot find module → run the install	knowledge
Structural classifier over hand-set field	stays — and the runner's pack-fingerprinting embodies it
Avoid default values	stays (judge)
git ls-files for tree-walking checks	stays; runner obeys
Comments carry why; a path spelled in one canonical place	check (adv), partial: same path re-spelled across many comments; the why half stays (judge)

skills/writing-tests/SKILL.md

Rule	Disposition
See a test fail before trusting it	stays
Run a new transform over the real corpus	stays
A stub at the bug's boundary can't catch it	stays
Teach a new call to every double	stays
Segment coverage gates by verification kind	stays
Drive goldens through the real code path	stays (judge)
Snapshot tests inject a fixed reference instant	check (adv): Date.now() / argless new Date() in snapshot/golden test files
e2e twice-green before merge	per-project CI policy; text stays for now
CI-only tests self-diagnose and are hang-proof	check (adv), partial: missing test-level timeouts
Hermetic self-authored inputs over live targets	stays
High-watermark gating for fuzzy metrics	stays (pattern; reusable pack helper later)
Human-reviewed expectation files	stays
Never silently re-baseline a golden	check (adv): diff changes a committed baseline artifact → finding demands the approval flow
Record the artifact first, read expectations off it	stays
Deterministic regeneration; never shell to git from a test	check (adv): git invocations inside test files

tasks/ — the rest

Doc	Disposition
bug-investigations.md (all 3 rules)	stays — pure investigative process → bug-investigation skill
agent-architecture.md (all 4 rules)	stays — it is this design's own rationale → folds into the unattended-agents skill
agenticBestPractices.md (~13 rules)	stays, mostly in-flight agent behavior → unattended-agents skill. Convertible slivers: @import-only-baseline → check basics/claude-md (adv) flagging new @-imports; per-routine tracking issue → check (adv), deferred (needs API)
agentic-documentation.md (~20 rules)	mostly stays (judge) → authoring-agent-docs skill. Convertible: <200-line budget → check basics/claude-md; emphasis ("IMPORTANT"/"YOU MUST") count threshold → check (adv); examples in <example> tags → check (adv)

(The stays-residues of testingPractices, textAndFileManipulation, and git-and-github likewise become the writing-tests, repo-text-sweeps, and git-github-advanced skills.)

technologies/

Doc	Disposition
aws-sam.md (3)	all three carry static signatures → check aws-sam/ pack: Handler still naming the entry's subdirectory under BuildMethod: esbuild; esbuild declared in devDependencies; a custom origin-request policy listing Authorization. Diagnostic residue → aws-sam skill
chrome-extension.md (8)	~4 have signatures → check chrome-extension/ pack: SetIcon({path: …}); non-root-absolute paths handed to chrome APIs in the service worker; interactive: false paired with prompt=consent; getAuthToken in a JWT-backend repo (adv, heuristic). Signature-less rest (shared-global augmentation, CDP probing, CORS) → chrome-extension skill
nodejs.md (2)	heuristic signatures only → (adv) lints in the node pack (innerText || textContent fallback in jsdom-tested code; jsdom fragment parsing without runScripts); knowledge → nodejs-testing skill
html.md (1), flutter.md (stub)	no static signature (runtime parser behavior) → skill
chrome-extension-release.md — the contract	check chrome-extension-release/ conformance suite: the single Release to Chrome Store workflow stub (file chrome-extension-release.yml, named Release to Chrome Store, daily schedule at the contract cron, calling the three canon reusable workflows) · no surviving __TOKENS__ · required .github/release.config with all 5 explicit keys, no defaults · manifest ↔ package.json version sync · zip forced to dist/<kebab repo>.zip (derived) · README Install/Releasing sections · PRIVACY.md present · every manifest permission disclosed in PRIVACY.md (test-the-world) · an added permission raises an advisory to open the store-dashboard justification issue (test-the-work) · release machinery under dev/build/release/ · store-asset inventory (128px icon, ≥1 screenshot) · assets generator-backed. The patch-bumper and shipped-paths filter are now canon composite-action/workflow logic, not per-repo scripts. Listing copy + permission justifications are Chrome Web Store dashboard state (no repo file); first publication is a one-time tracking issue, not a STORE-LISTING.md; no per-repo releasing.md. Doc keeps the setup narrative and manual store steps

templates/ & bootstrap.md

Rule	Disposition
Every project declares a category in its CLAUDE.md	check basics/claude-md (adv): consumer CLAUDE.md lacking a category/template declaration
Bootstrap steps, template catalog, generator prompt	stays — procedures run at adoption, not per session. Bootstrap gains the Stop-hook registration step (Phase 1)