Agent Skills: Living Documentation

Living Documentation

Knowledge sharing that stays correct with the least effort. Distilled from Cyrille Martraire, Living Documentation: Continuous Knowledge Sharing by Design (Addison-Wesley, 2019) - cite the book as the canonical source.

The whole method hangs off one maxim: no mechanism, no trust. A durable fact without something keeping it true is rot-in-waiting, and one stale fact a reader catches destroys the authority of every other fact in the document.

Use when

• Writing or reviewing an AGENTS.md / CLAUDE.md, a SKILL.md, an ADR, a README, or a knowledge-base note.
• Deciding whether a thing is even worth documenting.
• A doc keeps drifting from reality, or feels too long to trust.
• Designing how a project's knowledge should be organised.

The codex (sticky maxims)

Lead with these; they are meant to be memorable, not nuanced. Depth is in the references.

1. No mechanism, no trust - every durable fact needs an accuracy check, or it is rot-in-waiting.
2. Document only the delta from defaults - say only what a competent reader does not already know.
3. Rule of Two - write the rule the second time you correct someone, never speculatively.
4. Code shows what and how; documentation exists for why.
5. The best documentation never has to be read - it fires at the right moment, or makes the wrong thing impossible.
6. Put knowledge on the thing it describes - so it moves, renames, and dies with that thing.
7. Date it and let it be - accounts from the past need no upkeep; only current-state docs do.
8. Reference volatile -> stable, never the reverse - churn must not cascade upward.
9. One document, one message.
10. If it is hard to document, fix the design, not the prose.
11. Link, don't re-explain - name the canonical source and write only your 1%.
12. Derive it; never store the derived thing as a source.

The eight themes (map)

Each maps to a section of references/principles.md (the general method) and references/agent-docs.md (applied to AGENTS.md / skills / ADRs / KBs).

1. Accuracy & trust - pair every fact with a mechanism; the hierarchy of accuracy; dated accounts need none.
2. What earns a place - the default is don't; salience, Rule of Two, sedimentation, link-don't-re-explain, biodegradable docs.
3. Where knowledge lives - co-location, evergreen vs volatile, volatile->stable references, perennial naming.
4. Rationale & decisions - record the why and the rejected alternatives; commit messages as docs; the architecture codex.
5. Enforcement over prose - turn rules into types/linters/tests/hooks; the config is the doc.
6. Curation & navigation - one message per doc; highlight the core; sightseeing maps; make it skimmable and searchable.
7. Generated / living artifacts - consolidate dispersed facts; living glossary; exploit knowledge already in tools.
8. Feedback loops - docs as a design mirror; the cold-newcomer astonishment report; the two-minute test.

How to use this skill

• Authoring / reviewing a doc? Run the relevant checks from references/rituals.md: the salience pass, the two-minute test, evergreen/volatile classification, and (for any hand-maintained fact) "name the mechanism that keeps this true".
• Doc keeps rotting? Reach for theme 1 + 5 + 7: add a reconciliation check, enforce the rule instead of stating it, or derive the fact from its source.
• Deciding what to capture? Apply theme 2: Rule of Two, sedimentation, and the three gates (long-lived? many readers? critical?).

Don't duplicate sibling skills

Per maxim 11, point at these rather than restating them:

• mechanical-enforcement - which rules to turn into linters/hooks and how (theme 5).
• architecture - domain modelling, ports/adapters, making illegal states unrepresentable (theme 5.3).
• testing / test-coverage - reconciliation tests, contract tests, enforcement gates (themes 1, 5).
• hk - wiring pre-commit hooks and local checks for the mechanisms above.