docs-instrument

A repository has adopted the docs-quality rule set. Most of its rules carry a runnable verification, and none of those verifications runs yet. This skill turns that rule set into a gate that actually fails.

Work in cost order. The declaration retrofit is first because eight other families read the declaration, and none of their checks can classify a page until it lands.

Stop condition

Stop when all five hold. Do not keep going, and do not add rules.

1. Every checks/ script the repo wires up exits 0 with --self-test.
2. Every wired check has been seen red once, on a planted violation you reverted.
3. Every file path named in a runner target, a lint config or a rule row resolves on disk.
4. The signal manifest names every signal, including the deferred ones and their preconditions.
5. The whole-tree warning counts are recorded as the ratchet baseline.

A check that cannot go red is not coverage. A path that does not resolve is not a check. Both are the failure this whole gate exists to prevent.

Before you start

• Read rules/docs-quality/checks.md. It carries one row per script: what it checks, how to run it, its exit codes and its fixtures. Every flag beyond the common four lives there. Never invent a script name or a flag.
• Every Python check shares one interface: python3 checks/<name>.py [--root DIR] [PATH ...] [--format text|json] [--self-test]. Exit 0 is clean, 1 is findings, 2 is a usage error or missing input. Findings print as path:line: DOC-XXX-nn: message.
• Find the generator config: mkdocs.yml, book.toml, .vitepress/config.*, docusaurus.config.*, astro.config.*, docs/conf.py. The answer decides most of the wiring. See references/wiring-by-generator.md.
• Do not rewrite prose in this skill's work. Retrofits and configs only. A content fix belongs in its own change.

The procedure

Step 1. Build the published-file list

Every check reads this list. Getting it wrong is how a prose lint ends up reporting on research notes and agent config (DOC-PLAIN-23).

The list is git ls-files under the directory holding the generator config, plus repo-root README.md and CHANGELOG.md. Assert it excludes .agents, .claude, .serena, .worktrees, node_modules, dist, target and any build output directory.

A repository with a committed docs tree and no generator config still needs a list. Name that directory once, in the runner. The DOC-NAV family then reports not applicable and exits 0, because a tree with no site carries no sidebar and no search box (DOC-NAV-01).

Output. One shell function or task variable in the runner that prints the list. Schema in references/rollout.md.

Step 2. Retrofit the page declaration

checks/doc_declaration.py reads a doc_type and doc_tier comment inside the first 12 lines, using the comment opener of the file's markup family. It never reads a path (DOC-TYPE-02). Expect a 100 percent failure rate on the first run.

1. Run python3 checks/doc_declaration.py --root . --seed. It proposes a type per page from the nav config or the page heading. Nav-label seeding measured 94.3 percent accurate over 122 pages (wave2-declaration-key.md section 10), against 68.1 percent for a path classifier.
2. Review every seeded value. Read the content for the pages the seed could not place, which is roughly a third of a flat tree.
3. Add doc_tier on tutorial, how-to and landing pages only. The other six types take a type line alone (DOC-DISC-13).
4. Land it as one commit that changes nothing but declaration lines.

Traps that destroy files, all measured (wave2-declaration-key.md):

• Never write the declaration as YAML frontmatter. mdBook renders it as a visible fake heading and indexes it (DOC-TYPE-28).
• Never place the comment above existing frontmatter. It destroys the frontmatter on all three tested generators (DOC-TYPE-29).
• Use {/* */} in MDX. An HTML comment is a hard build error there, and Docusaurus applies MDX parsing to plain .md unless markdown.format is set to detect (DOC-TYPE-30).

Output. One retrofit commit, and checks/doc_declaration.py exiting 0 over the file list. Commit shape in references/rollout.md.

Step 3. Wire the checks into the existing runner

Use the task runner or test runner the repo already has. Do not add a second one.

Wire the scripts the rule set names: prose.py, page_type.py, landing_check.py, nav_depth.py, doc_examples.py, links_raw.py. They all import strip_prose.py, which strips front matter, declaration comments, fenced code, code spans, link targets, images, tables, HTML, admonition markers and include directives. Never hand a check raw file text.

Every one of them runs twice, and this is the rule that decides whether the gate survives its first week:

• Once over the changed files, at error severity.
• Once over the whole list, at warning severity, until the backfill lands.

A check may launch at error whole-tree only at zero standing violations. The median adopting page already fails several prose rules, so a whole-tree red gate blocks every open pull request on day one. Record each warning count as the ratchet baseline. Full schedule in references/rollout.md.

Output. Two runner targets, docs:check at error on the diff and docs:audit at warning whole-tree, plus a recorded baseline count per check.

Step 4. Add markdownlint, and Vale only as a second tier

Tier 0 is grep and markdownlint and is always present. Tier 1 is Vale, only when the repo accepts a new binary. No rule may take its only verification from tier 1, because a Vale-only check is an unchecked rule everywhere Vale is absent.

Use the shipped checks/markdownlint.jsonc. It sets MD025 with front_matter_title empty. The default treats a frontmatter title: as a second H1, which is a 100 percent false-positive rate against that convention. MD054 takes per-style booleans and has no style key. MD041 is off wherever the declaration comment is line 1.

Give every checkable construct exactly one owning tool, and disable the other tool's equivalent (DOC-PLAIN-20). Link style is the known collision. markdownlint MD054 owns it, so the Vale rule that demands the opposite stays off.

If Vale goes in, use checks/vale.ini. Pin each package by org and repo, never by display name, and print the resolved rule count and error share before enabling it. Two separately authored packages ship under the same popular name, one with 17 tiered rules and one with about 111 rules all at error.

Output. A markdownlint target on the changed-file list, and optionally a Vale target that no rule depends on alone.

Step 5. Wire link checking

Two passes, and they answer different questions (DOC-OBS-01, DOC-OBS-02).

Built output. The generator's own strict build satisfies this. Seven of nine measured sites already pass through mkdocs build --strict rather than through a separate checker (wave2-calibration-a.md section 3). Do not replace a working strict build with lychee. Where no strict mode exists, run lychee --include-fragments <build-dir> after the build, with <build-dir> taken from the generator row in references/wiring-by-generator.md.

Raw markdown. Any pre-build pass needs a source root and an exclusion for every page whose anchors are generated at build time. Without both it either floods the log or checks nothing. One measured repo produced 65 phantom dead links from a single four-line generated stub.

checks/links_raw.py resolves explicit {#id} anchors and root-relative paths, skips build-time-anchor pages, and lists what it skipped. checks/lychee.toml is the built-output example.

Prove it once. Break one anchor and one root-relative link in a fixture, confirm a non-zero exit, revert.

Output. One built-output gate and one configured raw pass, both seen red.

Step 6. Stand up the tested-example harness

Reach for the language's own doctest runner before writing anything (DOC-EX-03). The per-language table, with the current tools and their real limits, is references/tested-examples-by-language.md. Read it before you choose.

Where no runner fits, the harness is one file. It globs the example tree, runs each file as a subprocess and asserts the exit code (DOC-EX-04). The subprocess harness mode of checks/doc_examples.py is that file.

Bind a page to its test with a declared key in the test header, # doc: <slug>, never with a mirrored file path (DOC-EX-02). A mirrored path breaks every page the first time the test tree moves.

The gate itself is a set diff. Fences whose info string is on the runnable tier list, against fences carrying a binding key. A non-empty difference fails (DOC-EX-01).

Tag every fence from that tier list (DOC-EX-05). Write a tier suffix as one hyphen-joined token such as python-no-run. A space in a fence info string is unparsed under MkDocs Material and swallows the next fence (DOC-EX-20, DOC-EX-21). Wrap a snippet that must not run in a paired marker carrying a reason (DOC-EX-06).

The recording layer is optional and stays out of the gate (DOC-EX-11). Disable recording, re-run the required gate, and the result must not change. Every recording comes from a real run (DOC-EX-12), and a cast is committed only when no build step regenerates it (DOC-EX-13).

Output. A harness bound to the pages, in the same required gate as the unit tests. A failing example names the doc page, not only the test file (DOC-EX-07).

Step 7. Add the reader signals the stack can carry

Add what the stack supports today, and defer the rest with its precondition named (DOC-OBS-12). A requirement no repo can satisfy trains readers to ignore the whole rule set.

In ascending cost:

Sinks, vendor comparison, the cost bands and the deferral form are in references/reader-signals.md. The per-generator beacon wiring is in references/wiring-by-generator.md.

Two things to refuse. Enabling GA4 Enhanced Measurement never satisfies the zero-result requirement. Its view_search_results event triggers on one of five URL query parameters, and an overlay search writes none. Never bind the beacon to a generator's internal search file either (DOC-NAV-16).

State the review trigger as a count or a release boundary, never as a bare cadence word (DOC-NAV-15).

Output. docs/.meta/observability.md, holding every signal with a status, a review trigger and a bias disclosure (DOC-OBS-10). Schema in references/reader-signals.md.

Step 8. Add a Lighthouse ratchet where a built site exists

Only where the repo already builds a site. Assert the measured category scores, a point or two below the current median, and raise the floor as it improves. The one real measured instance runs accessibility at 0.97 against a median of 1.00. Best practices sits at 0.93 against 0.96 and SEO at 0.97 against 1.00. Performance is a warning at 0.85 against a median of 0.88.

Prove the gate red once. A missing alt, an empty <button> and an unlabelled <input> moved one real site from 0.92 to 0.77 and failed the gate. Record that proof beside the config. Schema in references/rollout.md.

Step 9. Every new check the adopter writes

This applies to a check you write, not to the shipped ones.

1. Write a fail-*.md fixture the check must reject, and a pass-*.md it must accept, under checks/fixtures/<script>/. A check with no rejecting fixture has never been proven able to fail.
2. Run the check over the real corpus and record its hit count and its false-positive rate.
3. Only then choose a severity. A check stays at warning until that rate is measured, and no unmeasured check goes above warning.
4. Every number in the finding message names its source in parentheses.

Fixture and measurement shapes are in references/rollout.md.

Run these before you call it done

# every shipped check proves itself against its own fixtures
for s in checks/*.py; do python3 "$s" --self-test || echo "SELF-TEST FAILED $s"; done

# every path any rule or config names resolves on disk
grep -ohE 'checks/[A-Za-z0-9_.-]+' rules/docs-quality*.md rules/docs-quality/*.md \
  | sort -u | while read -r p; do test -e "$p" || echo "MISSING $p"; done

# no unwired gate is described as coverage
grep -rniE 'tracked, not built|future gate|not implemented yet|(gate|check).{0,15}TODO' \
  rules/ .github/ 2>/dev/null

# no absolute or repository-specific path leaked into a template or fixture
grep -rnE '/(home|Users)/|(crates|services|packages)/' checks/ | grep -v Binary

Then plant one violation per wired check, watch it go red, and revert it.

Failure modes an agent falls into here

Ranked by how often the measurements caught them.

1. Names a check script and never writes the file. The dominant defect in this domain, measured at 17 rules and 7 phantom files. Verify with the path-resolution grep above.
2. Ships a checker that checks nothing. Points lychee at the raw markdown tree with no source root and no fragment flag. It sees zero issues and reports link checking as done.
3. Writes a check that cannot fail. A count compared to itself. A grep on a beacon with no listener. A probe that reports "cannot verify" and passes.
4. Launches the new lint at error across the whole tree because strict feels safer, and blocks every open pull request.
5. Classifies pages by path because the directory name looks right. A path classifier is wrong about a third of the time, and a path glob for runbooks matched 0 of 248 real pages.
6. Parses a nav config with a strict YAML loader. yaml.safe_load hard-fails on 4 of 7 real configs over !ENV and !!python/name: tags.
7. Assigns a severity from how important the rule feels, before the check has run once against real content.
8. Copies a tool's own space-separated fence attribute into the tier scheme, which silently eats the next fence under MkDocs Material.
9. Builds a bespoke harness for work an installed doctest runner already does, then sizes it like the large worked example it read.
10. Fires the nav and search family on a bare docs tree with no generator. Every finding is false on a repo that was never a docs site.
11. Proposes analytics as zero-result capture, missing that the automatic detection needs a URL parameter an overlay search never sets.
12. Adds a feedback widget first because it is the most visible ask, with no traffic denominator behind the percentage it will print.
13. Invents a freshness interval such as "review every 90 days". No source validates one, and a date stamp is never a gate (DOC-OBS-13).
14. Leaks a source path into a portable template, usually by filling the trigger matrix with the rows it just read.
15. Wires a rewrite-in-place example runner into CI and calls it tested. That tool accepts a broken command's new output as the new expectation (DOC-EX-23).

References

Read one level down, on demand. These files do not link each other.