docs-coverage

Docs coverage

A skill for auditing whether a design system's documentation surface keeps pace with its components. It measures the code (the source of truth for what exists) against each documentation surface and reports three things: coverage gaps (components with no documentation), staleness (documentation that predates the component's last code change), and orphaned docs (pages for components that no longer exist). Produces a severity-rated finding table with per-signal confidence labelling.

Before you begin: verify references

Confirm that every path in this skill's frontmatter references: exists relative to this SKILL.md. If any is missing, stop: the install is incomplete, usually because a flattening installer (for example npx skills install) dropped the repo-root knowledge-notes/ directory. Tell the user to reinstall by a method in 1-INSTALL.md and run verify-install.sh from the install root. Proceed without the references only if the user explicitly says to, and then say in the output that it was produced without the pack's reference material.

Context

Code is the source of truth for what components exist; the documentation surface is measured against it. A component in code with no docs is a coverage gap; a page for a deleted component is an orphan; a page older than the component's last change is a staleness risk.

This skill is built to work with no integration at all: a components directory plus a Storybook build plus git history answer coverage and staleness for most teams. Hosted platforms (Zeroheight, Supernova, custom docs sites) are optional layers that light up when configured — never prerequisites. The audit never blocks on an integration; it logs what is unavailable and proceeds with what it can reach.

The hard part is trust. Coverage is a join between two lists — components in code and entries in the doc surface — and the join is only as reliable as the key that links them. Every coverage finding carries a confidence tier so a fuzzy guess is never presented as a fact. See documentation-coverage.md for the full model.

Configuration

If .ds-ops-config.yml exists, follow the configuration-and-recurring knowledge note (../../knowledge-notes/configuration-and-recurring.md) for loading, integration fallbacks and recurring runs. This skill reads:

system.framework — affects how component files are discovered (e.g. .tsx / .vue / .twig)
severity.* — severity-rating overrides
integrations.storybook.static_path — local Storybook build directory (e.g. storybook-static); the preferred source
integrations.storybook.url — published Storybook URL for pulling /index.json when no local build exists
integrations.documentation — optional hosted platform: platform, url, api_key_env, plus styleguide_id (Zeroheight) or design_system_id (Supernova)
integrations.github — change history when the audit runs outside a local clone
docs_coverage.staleness_threshold_days — grace window before a doc is flagged stale (default 90)
recurring.* — trend comparison (see Recurring workflow)

Auto-pull integrations

Storybook — the primary surface (integrations.storybook.enabled: true or a local build):

Prefer a local static build: read the index from integrations.storybook.static_path (default storybook-static/index.json). No server, no auth.
If only a URL is configured, fetch <url>/index.json.
No build and no URL: read the CSF files directly. *.stories.* grouped by their title (or the default export's component) give rung 1; a tags: ['autodocs'] entry or a sibling MDX file gives a docs page. Say under Scope that the inventory came from story source, not an index, and that componentPath joins weren't available.
Branch on the top-level v field, which tracks the Storybook version: v: 3 is the SB 6 stories.json (entries under the stories key); v: 4 and v: 5 are the SB 7+ index.json (entries under entries), and v: 5 (SB 8.1+) adds componentPath. componentPath is opt-in and not guaranteed even on recent Storybook — use it for the Tier A join when present, and fall back to the Tier B name join whenever it is absent, regardless of v. A v: 3 stories.json has no type: 'docs' entries at all, so it can't show rung 2: mark rung 2 unknown for v3, unless you read the MDX files or each story's parameters.docs directly.
The official Storybook MCP (@storybook/addon-mcp) is optional — at the time of writing it needs a running server and is React-only/experimental. Use it only if the tools are already available; never make it a dependency.

Documentation platform — optional layers (integrations.documentation.enabled: true):

zeroheight: use the REST API (Enterprise) — GET /styleguides/{id}/pages and GET /pages/{id} give the documented-page set and updated_at per page (high-confidence staleness). There is no components endpoint, so reconstruct the coverage diff by name-matching pages against the code/Storybook inventory (Tier C).
supernova: use the MCP "Relay" or @supernovaio/sdk — get_design_system_component_list vs get_documentation_page_list gives a coverage diff (heuristic link, Tier B/C). Per-page staleness timestamps are not reliably exposed — mark staleness unknown unless a page timestamp is actually present.
custom: crawl the sitemap or rendered HTML for page titles; name-match only (Tier C).

Step 0: Build the component inventory (the source of truth)

Establish what components exist before looking at any documentation.

If a codebase-index output exists (.ai/index/), use its component list — it is already resolved and classified.
Otherwise, if the package has a public barrel (src/index.ts, or each package's entry point in a monorepo), use its component exports as the inventory — that's what consumers see and what needs docs. Internal building blocks that aren't exported don't need public docs; count them separately if at all.
Only if there's no barrel, glob the components directory. Common roots: src/components/**, packages/*/src/**, lib/components/**, app/components/**. Treat each component file/symbol as a candidate (e.g. Button.tsx, not Button.test.tsx, Button.stories.tsx, or index.ts barrels), and say the inventory may include internal components.
Record, per component: name, resolved file path (repo-relative, forward-slash), and category if a classification is available.

Produce a brief inventory line before continuing: N components found across M directories. If discovery finds nothing, ask the user where components live.

Step 1: Build the documentation inventory

For each available surface, list what is documented.

Storybook: parse index.json. Group entries by title. For each component, record: has a type: 'story' entry (rung 1, exists); has a docs page — autodocs or MDX via tags (unknown for a v3 stories.json, see above); and the resolved componentPath and importPath. A docs page alone doesn't make a component described: the autodocs tag generates a page even when no prop has a description. Rung 2 needs the page and descriptions on most props, from argTypes in the CSF file or JSDoc on the props interface. Record the prop-description ratio per component (described props of total) and report it; a page with 0 of 12 props described is rung 1 with a docs page, not rung 2. Also note play presence (interaction tests) where the CSF is read.
Figma component descriptions (Console MCP figma_get_component, or the official MCP's get_design_context on the library node): a documentation surface for designers. Record which components have a non-empty description and report it as its own column, not merged into the Storybook rungs.
Hosted platform (if configured): list documented pages and, where exposed, their updated_at/last-modified timestamp.
Usage guidance (rung 3): if the system documents usage separately (a usage-guidelines output, MDX "When to use" sections, a Zeroheight guideline page), record which components reach rung 3.

Step 2: Join the inventories

Join code components to documentation entries using the Tier A / B / C reliability hierarchy in documentation-coverage.md (file path, then symbol name, then fuzzy title match), and record the tier on every match. Never state a Tier C result as fact.

Before reporting anything as undocumented, run a positive control: confirm the join matches a component you can see is documented (open its story or page). If the join misses it, the key doesn't fit this codebase — fix it before reporting gaps.

Step 3: Coverage gap analysis

From the join, produce:

Undocumented components — in code, no entry on any surface. Report each with the join tier that found (or failed to find) it. Group by the documentation rung they fall short of.
Rung distribution — how many components reach exists / described / guided. A high story count with few described/guided components is itself a finding.
Orphaned docs — pages/stories whose component no longer exists in code. These point the opposite direction and are usually quick removals. Before calling a page orphaned, search the codebase for the name it documents (including renamed or re-exported symbols) — a Tier B/C "orphan" is often just a title that drifted from the component name.

Step 4: Staleness analysis

For each documented component, compare change dates:

Component last change: git log -1 --format=%cI -- <component source path>. If the latest commit looks test-only or cosmetic, check it with git log -1 --name-only -- <component path> and lower confidence rather than excluding file types.
Doc last change: git log -1 --format=%cI -- <story/MDX file>, or the platform timestamp (updated_at).
Flag stale when component_last_change − doc_last_change > staleness_threshold_days (default 90). A doc the same age as or newer than the component is never stale.
git log returns empty output (not an error) for a file with no commits in the current branch/clone. Treat an empty result as untracked and mark staleness unknown — never as fresh. Do the same when a platform exposes no timestamp. Prefer a full clone: shallow clones and renames without --follow give misleading dates.

See documentation-coverage.md for why this proxy over-flags, how staleness and join confidence combine, and how to frame stale findings as a risk ("docs predate a code change on [date] — confirm they still match", with both dates shown) rather than a defect.

Step 5: Produce the report

Open with a headline sentence stating overall state and where to focus. Example: "Of 84 components, 71 have a story but only 38 have a docs page, and 9 docs pages predate a code change. The coverage floor is solid; the described/guided layer and 9 staleness risks are where to focus."

Structure the report:

Docs coverage report

Date: [date] Inventory: [N components] · Surfaces audited: [Storybook / Zeroheight / Supernova / custom] · Join confidence: [the tier used for most components — A, B, or C]

Summary One paragraph. Overall state, the most urgent gap, and an explicit note on which signals were measured vs estimated or unavailable. Write it like a peer review, not a compliance filing.

Coverage by rung — count each component at the highest rung it reaches. The rows are mutually exclusive (a "guided" component is not also counted under "described" or "exists"), so they sum to the full inventory. "Undocumented" is the rung-0 bucket: components on none of the three documentation rungs.

Rung	Count	% of inventory
Guided (usage guidance)
Described (docs/autodocs page, no usage guidance)
Exists (≥1 story, no docs page)
Undocumented (no surface)

Findings List each finding with:

Finding ID (e.g. DC-01)
Severity: 🔴 Critical / 🟠 High / 🟡 Medium / ⚪ Low
- 🔴 Critical — a foundational component (Button, Input, Text, Icon, or anything with high fan-in) with no documentation on any surface (rung 0)
- 🟠 High — any other public component at rung 0; a foundational component stale on a high-confidence timestamp; any component stale on a high-confidence timestamp where the code change touched its props interface or rendered element (the doc is now wrong, not just old)
- 🟡 Medium — a public component stuck at rung 1 (a story but no docs page), or a stale doc on a lower-confidence timestamp
- ⚪ Low — orphaned docs, internal components, and Tier C gaps awaiting confirmation
- Only weight by traffic or usage if adoption data exists (see the adoption-measurement note); otherwise use foundational status, which the code can show
Category: Coverage gap / Staleness / Orphan
Confidence: Tier A / B / C (and timestamp source for staleness)
Evidence: the component's file path and the docs entry it was joined to (story id, docs page URL) or the surfaces searched when none was found; for staleness, both change dates with their sources
Description, affected component(s), and recommended action

Orphaned documentation List pages/stories with no matching component, with the suggested removal.

Action list

Immediate: undocumented foundational components (or high-usage ones, where adoption data exists); orphans
Planned: raising the described/guided layer; resolving staleness risks
Review: Tier C matches needing manual confirmation

Scope

Inspected: [inventory source (barrel, index, or glob), Storybook index version, platforms queried, git history depth]
Not inspected: [surfaces out of reach, e.g. a docs platform without API access]
How "none found" was checked: [the join's positive control, and which rungs were unknown rather than absent — omit if the report makes no absence claims]
Assumptions: [anything taken as given rather than verified]

End with the closing note below.

Recurring workflow

Follows the recurring-run procedure in the configuration-and-recurring note. Specific to this skill:

Compare coverage-by-rung deltas, newly undocumented components (regressions — these are the priority), newly resolved gaps, and staleness count trend.
Add a "Trend since last audit" section to the header: previous date, rung deltas, and one sentence — "Documentation coverage is improving / stable / declining since [date]."

Closing note (include in every report)

End the report with:

A note on context: This audit measures your documentation surface against your code — it does not see why a component was left undocumented or why a doc predates a change. Some gaps are deliberate (internal-only components) and some "stale" docs are still correct after a refactor. Tier C matches are best-guesses, not facts. If a finding describes an intentional choice, let me know — I'll calibrate future runs. The goal is to surface drift you haven't seen, not to second-guess decisions you've already made.

Quality checks

Every coverage finding carries a join confidence tier; no Tier C result is stated as fact
Staleness findings show both change dates and name the timestamp source; unavailable timestamps are marked unknown, not assumed fresh
The report distinguishes the rungs (exists / described / guided, plus the undocumented rung-0 bucket), counted at highest-attained — not a single coverage percentage; "described" rests on prop descriptions, not on the presence of a docs page
Orphaned docs are reported separately from coverage gaps — they point the opposite direction
The summary states which signals were measured vs estimated or unavailable
The audit ran on whatever was available and did not block on a missing integration
"Undocumented" claims rest on a join that passed its positive control
The Scope block and the closing note about intentional choices are present