docs-drift-guard

Documentation drift guard

Scope

This skill owns documentation method. The repo owns protected documents, generated outputs, check entrypoints and review requirements. A working link establishes path integrity, not that the surrounding claim is still true.

Prerequisites

Read AGENTS.md and the relevant doc's source/ownership rules. Identify the facts being described, their code/configuration source and existing tests/generators. Use harness-operations only when tool mapping is needed. Do not introduce a new dependency merely to run this skill.

Procedure

1. Give each fact one authored home and link other consumers to it. Keep stable architecture intent in prose; derive API, CLI, configuration and mutable inventories from their owning source when possible.
2. Prefer generated output with freshness checks, then executed examples, then deterministic structural checks, then explicit review of remaining prose. A generated artifact still needs a check against its source; generation alone does not guarantee freshness. Keep model-based semantic review advisory.
3. Reuse the repo's existing link/path, doctest, snippet or architecture gate. If a new checker is needed, add it through the repo's dependency/CI process with a pinned version and prove it catches a broken reference or stale sample.
4. For JavaScript repos adopting the maintained docs-drift-check tool, use its locked local binary via the owning package script or npx --no-install docs-drift-check. Do not rely on an unpinned runtime download. External-link checks are a separate opt-in network axis, not an offline requirement.
5. Match higher checks to the claim: doctests for runnable examples; deterministic transclusion/freshness checks for code snippets; dependency rules for layering; path-triggered owner review for prose that cannot be mechanically derived.
6. Edit the canonical source and regenerate outputs. Preserve ADR history through the repository's amendment convention. Do not make separate hand-edited copies of generated docs, common prompts or native skill projections.

Completion

Report authored source, generated consumers, checks/negative controls and their results. State which semantic claims still depend on review and which network checks were unrun. A file that exists can still describe obsolete behavior.