Living Docs

Treat documentation as a living system that stays in sync with the code, not a write-once artifact that rots. The discipline has one spine — every piece of knowledge has exactly one home, that home is indexed, and nothing structural ships without its doc — and several document types that hang off it: a constitution, ADRs, BDRs, PRDs, issues, research, architecture diagrams, and a semantic context index.

This skill is stack-agnostic. It governs how docs are organized and maintained, never what technology a project uses.

Core invariants (the spine)

These hold across every document type. Everything else is detail.

Docs-first. Author the body in the repo (docs/…) before publishing anywhere external (tracker, wiki). The repo file is the source of truth; the external copy is a mirror.
One home per fact. Each concept, decision, or requirement lives in exactly one file. No duplication — cross-reference instead of copying. Duplicated prose is drift waiting to happen.
Indexed or it doesn't exist. Every doc is reachable from an index (an index.md listing in its directory, and the bundle-root docs/index.md that the project guide links). No orphan files.
Supersede, never rewrite history. Decisions and requirements are append-only records. When something changes, mark the old record superseded and write a new one — never silently edit the past.
No structural change without its doc. New module, moved files, schema change, new data flow → update the relevant doc and its diagram in the same change. No "I'll document it later."

When in doubt, re-derive the right action from these five. The rules files below are just these invariants applied to each document type.

Format: OKF-conformant

The five invariants govern organization and lifecycle; the Open Knowledge Format governs the file format so the corpus stays portable and agent-parseable. Every doc in the system is also an OKF concept. Load the okf-knowledge-format skill (it vendors the spec) when authoring or checking format.

OKF is a thin, swappable dependency — not a foundation (version risk). OKF is v0.1 from a single vendor; a backward-incompatible v0.2 is a real possibility. The invariants above do not depend on OKF — they depend only on a small set of frontmatter fields, so an OKF break cannot take the governance layer down with it. Keep the boundary explicit:

Required by Living Docs (the fact contract scripts/lint-docs.sh enforces): a non-empty type, and status + superseded_by on superseded records. These are ours; they survive regardless of OKF.

Inherited from OKF (format conventions): reserved index.md/log.md, the bundle-root okf_version, bundle-relative links, the # References heading (§8). If OKF changes, only this row moves — re-pin the version in the okf-knowledge-format skill and adjust.

Two rules apply to every concept file:

Frontmatter with a required type. Every non-reserved .md doc opens with a YAML frontmatter block whose type names the doc kind (Constitution, PRD, ADR, BDR, Issue, Context, Architecture View, Research, Reference). Recommended: title, description, tags, timestamp. Living-docs adds producer keys: status, supersedes, superseded_by. Status moves into frontmatter — no **Status:** body line.
Reserved files + bundle-relative links. The bundle root is docs/. Directory listings are index.md (OKF §6, no frontmatter — except the bundle-root docs/index.md, which carries okf_version: "0.1"). Optional log.md records directory history (§7). Cross-link with /-prefixed bundle-relative paths (/adr/0007-slug.md); list sources under a # References heading (§8), each entry formatted per rules/citation-conventions.md — ABNT NBR 6023 structure, always carrying the link, with connective labels in the project doc language (default English: Available at: <URL>. Accessed on: <date>) per rules/doc-language.md.

Doc trail

Every change follows this chain, from foundational source of truth down to code:

flowchart LR
  C[constitution] --> P[PRD]
  P --> A[ADR]
  P --> B[BDR]
  A --> I[issues]
  B --> I
  I --> K[code]

Artifact	Role
constitution	Foundational source of truth: what the product is, core data model, non-negotiables. All other docs sit under it.
PRD	What the system must do and why — feature/product requirement spec.
ADR	How the system is structured — architectural/implementation decision and rationale.
BDR	What the system must observably do — inputs, outputs, side effects, Given/When/Then scenarios — and how each is tested (the Test Design matrix; single home for "how to test", an execution issue links it).
issues	Execution slices — discrete units of work that trace back to ADRs/BDRs.
code	Implementation — every behavior, structure, and interface specified above, realized.

When to invoke

Standing up documentation for a project (creating docs/ structure, the docs index, ADR/issue/BDR/constitution directories).
First time living-docs runs in a project (no ## Living Docs block in the project guide) → ask the enforcement-mode question and persist the answer → see Enforcement modes.
Adopting living-docs in an existing/brownfield project (decisions already made but undocumented) → follow Procedure → Adopting living docs in an existing project: inventory the decisions, confirm each with the user before recording any ADR, never back-fill by inference alone.
Writing or editing an ADR (an architectural/implementation decision) → load rules/adr-conventions.md + templates/adr.md.
Writing or editing a PRD (a product/feature requirement spec) → load rules/prd-conventions.md + templates/prd.md.
Writing or editing a BDR (observable behavior — inputs, outputs, Given/When/Then scenarios, and the Test Design matrix for how each is tested) → load rules/bdr-conventions.md + templates/bdr.md. A test-strategy decision (non-default level/technique, bar deviation) is an ADR tags: [testing], not a new record type (no "TDR").
Specifying a non-functional requirement (performance, availability, security, scale) → a quality-attribute scenario bound to an instrument in the PRD (rules/prd-conventions.md rule 9 + templates/prd.md); the decision + fitness function go in an ADR. Not a new doc type.
Establishing or amending the constitution (foundational scope, data model, non-negotiables) → load rules/constitution-conventions.md + templates/constitution.md.
Creating or editing an issue/ticket → load rules/issue-workflow.md + templates/issue.md.
Recording research (technology evaluation, external trade-offs) → load the research-artifacts skill. It owns the OKF research-note format (single file per note, no per-research subfolder), the source discipline, and the research → decision → issue traceable chain, and links back here for the ADR/BDR/issue artifacts. Pairs with the deep-research skill.
Drawing or updating an architecture, data-flow, or tool-calling diagram → load rules/architecture-diagrams.md + templates/architecture-index.md.
Defining a term or acronym the docs use → add it to the glossary (docs/context/glossary.md), one home per term → load rules/glossary-conventions.md + templates/glossary.md.
A doc has grown too large or mixes concerns → split into a semantic index → load rules/semantic-index.md + templates/context-index.md.
Enforcing the no-drift maintenance rule after any structural change → load rules/maintenance-invariant.md.
Authoring or checking the OKF format of any doc (frontmatter type, reserved index.md/log.md, bundle-relative links, # References) → load the okf-knowledge-format skill.
Deciding which language the docs are written in (default English; user may override at session start and pin it) → load rules/doc-language.md.

Document map

Type	Lives in	Purpose	Mutability
Project guide	`CLAUDE.md` / `README.md` (root)	Entry point: scope, stack, docs index, mandatory workflows	Live — edit freely
Constitution	`docs/constitution.md`	Foundational source of truth: product scope, data model, non-negotiables	Append-only once ratified (amendment log)
Context index	`docs/context/index.md` + group files	Domain & module vocabulary, semantically grouped	Live — edit freely
Glossary	`docs/context/glossary.md`	Terms & acronyms defined once, in the doc language (acronym headwords as-is)	Live — edit freely
Architecture	`docs/architecture.md` or `docs/architecture/` + index	Living Mermaid diagrams: structure, data model, flows, tool-calling	Live — must match code
ADR	`docs/adr/NNNN-slug.md`	One architectural/implementation decision	Append-only (supersede)
BDR	`docs/bdr/NNNN-slug.md`	One observable-behavior decision	Append-only (supersede or amend)
PRD	`docs/prd/NNNN-slug.md`	One feature/product requirement spec	Append-only once accepted
Issue	`docs/issues/NNNN-slug.md`	Tracker mirror (body), one per ticket	Body editable; published copy follows
Research	`docs/research/NNNN-<slug>.md` (single file, no subfolder; sequential number leads, date in frontmatter `timestamp`; ends in `# References`)	External evidence with sourced claims	Append-only (evidence is dated)

Each directory carries its own index.md listing (OKF §6, no frontmatter). The project guide's "Docs index" links to the bundle-root docs/index.md. See rules/semantic-index.md for the indexing contract.

Procedure

Setting up living docs in a new project

Ask the enforcement-mode question (see Enforcement modes → First-run question) before anything else, since no preference is persisted yet. Record the answer as the ## Living Docs block in the project guide; default to strict if the user has no preference.
Create the project guide (CLAUDE.md or equivalent) with a Docs index section and a Maintenance rule section (copy the wording from rules/maintenance-invariant.md). Use templates/claude-hard-rules.md as the starting point for the project guide's hard-rules section — it already carries the ## Living Docs enforcement block; fill in the placeholders before committing.
Create docs/ with the directories the project needs (adr/, bdr/, issues/, prd/, research/, context/). Seed docs/constitution.md from templates/constitution.md. Add the bundle-root docs/index.md (carrying okf_version: "0.1"), and give each directory its own index.md listing from day one — even if near-empty.
Seed the context index (docs/context/index.md) with whatever domain vocabulary exists, and the glossary (docs/context/glossary.md) with the terms and acronyms the docs already assume (rules/glossary-conventions.md). Grow both as concepts are named.
Seed the architecture doc (docs/architecture.md) with the high-level Mermaid views the system already has. Promote to a docs/architecture/ directory + index once it grows (rules/architecture-diagrams.md).
Record any already-made decisions as ADRs so they are not re-litigated — but confirm each with the user before recording (see Adopting living docs in an existing project, steps 3–5); never back-fill an ADR by inference alone.

Adopting living docs in an existing project (brownfield)

An existing codebase already embodies decisions that were never written down. The failure mode here is the agent back-filling ADRs by inference and presenting them as settled — recording decisions the user was never asked to confirm. Adoption is therefore an elicitation exercise, not a transcription one.

Ask the enforcement-mode question (first-run) and persist the ## Living Docs block, exactly as for a new project.
Scaffold without deciding. Create docs/ + each directory's index.md, the bundle-root docs/index.md, and seed the glossary/context index from vocabulary already present in the code, the README, and the agent guides (CLAUDE.md / AGENTS.md). This is mechanical — no decisions are made here.
Read the existing context first, then inventory the decisions — as candidates, not records. Harvest what the project already carries: the code itself, plus the README, the agent guides (CLAUDE.md / AGENTS.md), package manifests, and any design notes or comments. From that, produce a list of the load-bearing decisions the project appears to embody (stack, boundaries, data model, key trade-offs). Do not write ADRs yet.
Present the inventory to the user and confirm each. For every candidate, state the inferred decision and the alternatives it appears to have ruled out, and ask the user to confirm, correct, or discard it — grill the load-bearing ones (grill-me if installed). The user owns the decision; the agent only surfaces what the code implies.
Record only the confirmed decisions as ADRs. These are origin records — they supersede nothing. Capture the chosen option and the rejected alternatives the user confirmed. A candidate the user discards, or one whose rationale nobody actually knows, is not invented into an ADR.
Seed the architecture doc with the high-level Mermaid views the system already has (rules/architecture-diagrams.md), then resume the Maintaining loop below.

Maintaining living docs (every task)

Before coding: read the ## Living Docs enforcement mode from the project guide (if the block is absent, this is the first run — ask the first-run question and persist the answer). Then read the relevant constitution, ADRs, BDRs, and the context index. Decisions there are not to be re-opened casually.
While working: if you name a new concept, add it to the context index; if you introduce a new term or acronym, define it once in the glossary. If you make a decision with a load-bearing rationale, grill it before recording it — surface the decision, ≥2 materially-distinct alternatives, and a recommendation to the user (run the grill-me companion if installed, else inline), then write an ADR capturing the chosen option and the rejected ones. If you specify observable behavior, write or amend a BDR. Never record a decision the user was not asked about (see Enforcement modes → Mode governs completion, not elicitation).
In the same change: update every doc the structural change touches — index rows, architecture diagrams, vocabulary. Run the maintenance checklist (rules/maintenance-invariant.md).
Never leave an index stale, an orphan file unlinked, a diagram contradicting the code, or a superseded decision silently edited.

Composition with other skills

Bundled in this repo — installed alongside living-docs:

okf-knowledge-format — the portable, exchange-oriented format standard (Open Knowledge Format: markdown + YAML frontmatter, required type, reserved index.md/log.md, bundle-relative links, conformance rules). Living-docs governs which docs exist and the no-drift discipline; OKF governs how a knowledge bundle's markdown and frontmatter are shaped. Author a shareable knowledge corpus — e.g. docs/context/ vocabulary or a docs/research/ catalog meant to be consumed by agents or exchanged across orgs — as an OKF bundle so its frontmatter and indexing are spec-conformant. The two compose; they do not overlap.
research-artifacts — owns the research-note format and discipline (a single-file OKF docs/research/NNNN-<slug>.md — sequential number leads, date in frontmatter — with a trailing # References, plus the cross-research docs/research/references.md roll-up, source rules, the research → decision → issue chain). Living-docs delegates all research authoring to it and consumes its output: an accepted recommendation becomes an ADR/BDR here, which spawns issues. The links between the two are bidirectional.

Optional companions — these are not included in this repo; living-docs works without them, and only composes with them if you happen to have them installed:

grill-me (Matt Pocock's skills, MIT — make install-pocock) — before writing a PRD or a load-bearing ADR, grill the design to surface the real decision and its alternatives. The grilling output becomes the ADR's Context/Consequences. The Maintaining loop (step 2) and the brownfield-adoption procedure invoke this step explicitly; without grill-me installed, do the lightweight inline version — state the decision, ≥2 materially-distinct alternatives, and a recommendation, and get the user's call before recording.
A research-gathering skill (e.g. a deep-research tool) gathers and cross-verifies the evidence that lands in docs/research/; research-artifacts defines how that evidence is formatted, indexed, and referenced from ADRs/PRDs.
A code-structure index (e.g. a codegraph tool) is the structural index of code; living-docs' context index is the human-readable index of concepts and the architecture diagrams are the visual index of structure. Keep all current.
An execution/review step downstream consumes what these docs plan: it implements the issues and verifies that the code honors the ADRs and BDRs, and that every specified observable behavior (Given/When/Then) is tested.

Enforcement modes

The five invariants above are always hard stops — they hold in every project regardless of mode (an orphan file or a silent rewrite is never acceptable). What the mode governs is the doc trail (constitution → PRD → ADR/BDR → issues → code): how strictly the agent refuses a structural or behavioral change that ships without its decision record.

The mode is chosen once, by the user, the first time living-docs runs in the project, and persisted in the project guide. The default is strict — the discipline is opt-out, not opt-in, because a flow nobody is forced to follow is the flow that rots first.

Mode	Doc trail	Agent behavior
`strict` (default)	Mandatory	Refuses to report a structural/behavioral task complete without its required doc (PRD/ADR/BDR/issue). Same hard-stop weight as the five invariants.
`guided`	Prompted	Pauses and asks the user before skipping a doc-trail step. The user may waive a step per task; the waiver is not remembered.
`lite`	Advisory	Only the five invariants are hard stops. The doc trail is recommended, never enforced (this is the pre-0.3 behavior).

Mode governs completion, not elicitation

A frequent misread is that strict means "the agent interrogates every decision." It does not. The mode governs completion enforcement — whether a structural/behavioral task may be reported done without its doc. It says nothing about how the decision inside that doc was reached.

Decision elicitation (grilling) is a separate, always-on concern, independent of mode. A load-bearing decision is never recorded from the agent's own inference alone: before writing an ADR/BDR, surface the decision, ≥2 materially-distinct alternatives, and a recommendation to the user, then record the chosen option and the rejected ones. Run the grill-me companion (see Composition) to drive that interrogation if it is installed; otherwise do the lightweight inline version. Never write an ADR for a decision the user was never asked about — in any mode. Mode changes whether you may ship without the doc; it never licenses inventing the decision the doc records.

First-run question

When living-docs is invoked in a project and no enforcement preference is yet persisted (no ## Living Docs block in the project guide), ask the user once, before doing the work:

This project hasn't set a living-docs enforcement mode yet. How strictly should the doc trail (PRD → ADR/BDR → issues) be enforced?

strict (recommended) — every structural/behavioral change must carry its doc; I'll refuse to call a task done without it.

guided — I'll ask before skipping a doc-trail step.

lite — only the five invariants are enforced; the doc trail is advice.

Persist the answer immediately in the project guide (CLAUDE.md, where it enters context at session start), then proceed:

## Living Docs
enforcement: strict   # strict | guided | lite
onboarded: <YYYY-MM-DD>

Absence of the block is the only first-run signal; presence of any valid enforcement value means onboarded — never ask again, just read it and apply it. To change modes later, the user edits the block.

Doc-trail enforcement is a judgement call (there is no sound oracle for "did this change need an ADR"), so it lives with the agent, not with lint-docs.sh. The mechanical invariants (frontmatter, indexing, links, supersede) are checked the same way in every mode.

Agent enforcement (refusal triggers)

The value of this skill is not the discipline — humans abandon "nothing structural without its doc" at the first deadline. The value is an agent that enforces it automatically. So the invariants are not advice; they are hard stops. When acting as the agent, do not report a docs-touching task as complete if any of these hold — fix it or surface it first:

Orphan. A new or moved concept file is not listed in its directory index.md (and that directory is not reachable from the bundle-root docs/index.md). Indexed or it doesn't exist.
Stale diagram. A structural change (schema, module layout, data flow, new component) landed without updating its Mermaid diagram in the same change.
Silent rewrite. A decision/requirement was edited in place instead of superseded — or a record is status: Superseded with no superseded_by.
Untyped doc. A non-reserved .md is missing frontmatter or a non-empty type; or an index.md/log.md carries frontmatter (except the bundle-root index.md).
Broken link. A bundle-relative (/…) or relative link points at a file that does not exist.
Duplicate home. The same fact now lives in two files (cross-reference instead).
Broken doc trail (mode-gated — see Enforcement modes). A structural change shipped without its ADR, or a behavioral change without its BDR. Under strict this is a blocked task — refuse it like an orphan. Under guided, pause and ask the user before proceeding. Under lite it is advisory only.

Triggers 1, 3, 4, 5 are mechanical — run scripts/lint-docs.sh (below) and treat a non-zero exit as a blocked task, not a warning. Triggers 2, 6, and 7 are semantic (no sound oracle) and stay a judgement call: inspect the diff before declaring done. Trigger 7 additionally depends on the project's enforcement mode.

lint-docs — the deterministic instrument

scripts/lint-docs.sh [docs/] mechanically validates invariants 2, 3, and 4 (the ones a machine checks better than prose): frontmatter/type, directory-index membership + root reachability, link resolution, and supersede integrity. A constraint without an instrument is a vibe — so the checkable invariants get a checker. Wire it into the project's quality gate / CI; a docs PR that fails it does not merge. It does not check docs-first mirroring or "one home per fact" semantics — those have no sound oracle and stay with the reviewer.

./scripts/lint-docs.sh docs          # lint the project's bundle; exit 1 on any violation

It delegates parsing to mature tools (so it is correct, not a hand-rolled markdown/YAML parser): lychee (link validity, all link forms, --offline), yq v4 (frontmatter YAML), and jq (lychee's JSON). Install all three; the script exits 2 with install pointers if any is missing. The OKF structural graph (index membership + reachability) is the script's own logic on top.

If you would rather not install the tools, run the checker via its self-contained image (needs only Docker): docker build -f Dockerfile.lint -t living-docs-lint . then docker run --rm -v "$PWD:/work" living-docs-lint docs.

A worked, lint-clean corpus lives in examples/linkly/ — copy its shapes.

Quality checks

Before considering a docs change complete. The frontmatter, indexing, link-resolution, and supersede items are enforced by scripts/lint-docs.sh — run it rather than eyeballing them; the rest are judgement:

Notes

Keep individual docs scannable. When a context, vocabulary, or architecture file passes ~200 lines or mixes concerns, split it via rules/semantic-index.md rather than letting it grow.
This skill describes conventions, not tooling. Numbering schemes (NNNN), tracker choice (GitHub Issues, Jira), and directory names can be adapted per project — the invariants must not.
Templates in templates/ are starting points. Trim sections that don't apply; never delete a section just to avoid filling it in if it's load-bearing.

Provenance — instrumentalization, not invention

Almost every doc type here is established prior art; this skill composes and enforces them, it does not invent them:

"Living documentation" is Cyrille Martraire's named methodology (Living Documentation, 2019) — we adopted the term.
ADR is Michael Nygard's (2011); the structured-markdown form is MADR; the supersede-don't-delete status convention is the adr-tools convention (Nat Pryce) — invariant 4 is that convention, not a new rule.
BDR is Specification by Example / Given-When-Then (Adzic 2011; North 2006) in an ADR-style record (recent third-party coinage, Zanzal 2026) — see rules/bdr-conventions.md.
The problem-first PRD and vertical-slice issue workflow are a kindred influence from Matt Pocock's to-prd / to-issues skills (github.com/mattpocock/skills, MIT); both instrumentalize older practice — product PRDs, user stories (Cohn; Beck), tracer-bullet vertical slices (Hunt & Thomas; Cockburn), and "make the change easy, then make the easy change" (Beck). Living-docs adds the OKF format, the docs-first mirror, and supersede governance.
The format is OKF, a published Google Cloud Platform standard (v0.1, 2026-06-12), not ours — see the okf-knowledge-format skill.
Neighbors we converge on: Diátaxis (Procida), arc42 (Starke; Hruschka), C4 (Brown), and docs-as-code (Write the Docs).

The honest claim is not invention — arc42 + ADR + C4 + glossary + docs-as-code is the informal stack of plenty of mature teams. What this skill adds is the explicit, agent-enforceable packaging of that stack: the governance invariants (supersede-never-delete + one-home-per-fact + indexed-or-doesn't-exist) carried in frontmatter as a fact contract and wired to a deterministic checker (scripts/lint-docs.sh). The prior-art research found no prior assembly of that exact governance layer — but the value on offer is the enforcement, not the novelty. Full citations: ../../references/prior-art-landscape.md.

ejklock/living-docs-skill

Ask in your favorite AI

說明文件