Use when writing or editing anything under docs/docs/ (mkdocs site), README.md, CONTRIBUTING.md, CHANGELOG.md, docstrings, commit messages, PR titles, or GitHub Release notes for Area Occupancy Detection — or when a code change needs an accompanying doc update and you're deciding what/where to write. Trigger words: "update the docs", "add a features page", "write a docstring", "release notes", "mkdocs nav", "stale doc", "CHANGELOG".
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Use when writing or editing anything under docs/docs/ (mkdocs site), README.md, CONTRIBUTING.md, CHANGELOG.md, docstrings, commit messages, PR titles, or GitHub Release notes for Area Occupancy Detection — or when a code change needs an accompanying doc update and you're deciding what/where to write. Trigger words: "update the docs", "add a features page", "write a docstring", "release notes", "mkdocs nav", "stale doc", "CHANGELOG".
AOD docs and writing
What this covers
The docs-of-record for Area Occupancy Detection: the mkdocs site under docs/docs/, README.md, CONTRIBUTING.md, docstring style, commit/PR title conventions, and how release notes actually get published. It tells you where a given fact belongs, what house style looks like (with quoted exemplars), which existing docs are known-stale landmines, and how to avoid re-creating the #491 stale-doc problem (code changed, doc didn't).
When NOT to use this
Deciding what the Bayesian math should say (formulas, constants, thresholds) — that's bayesian-occupancy-reference.
Deciding whether a code change is safe / needs a migration / breaks configs — that's aod-change-control.
Writing content for aod-external-positioning (README competitive framing, HACS listing copy) beyond the structural conventions below — check that skill for messaging strategy.
Investigating why a historical bug happened — that's aod-failure-archaeology.
Verify the live nav any time you add/remove/rename a page — mkdocs.yml's nav: block is a hand-maintained list, not auto-generated. A page that exists on disk but is missing from nav: will not appear in the built site (confirmed by reading docs/mkdocs.yml 2026-07-06: nav lists 14 features pages and 14 technical pages, one entry per file, no glob).
Doc-site build/deploy is docs.yml (GitHub Actions workflow, referenced by the docs badge in README) — it's the CI job that runs mkdocs build/gh-pages deploy; a stray gh-pages branch in the repo is its output, not something to hand-edit.
The features/ vs technical/ split
Rule of thumb: features/ answers "what do I get and how do I configure it"; technical/ answers "how does this actually work under the hood." Every features page should be readable by a user who has never opened the Python source. Every technical page assumes the reader is about to read or modify code.
The newest and best exemplar pair for this split is features/adjacent-areas.md (user outcome) paired with technical/transition-learning.md (mechanism). Use this pair as the template when writing a new feature's docs:
features/adjacent-areas.md opens with the user-visible effect ("Adjacent Areas lets rooms that are physically connected... influence each other's occupancy calculation"), covers configuration steps, an "Example" walkthrough in plain language, an "Observing it" section pointing at diagnostics, and an FAQ. It explicitly punts implementation detail: "Both effects come from learned transition history, not from a fixed 'influence' setting you configure. See Transition Learning for how that history is built and the underlying maths."
technical/transition-learning.md opens with the storage schema (AreaTransitions table columns), the detection algorithm (pipeline step name, deque-based windowing, exact constants like ADJACENCY_TRANSITION_WINDOW_S (60s)), and a 6-level smoothing-fallback table with threshold constants — content a features page would never include.
When you write a new feature, produce both files as a pair, cross-linked in both directions, rather than one page that tries to do both jobs.
House style
Plain prose over bureaucratic tone; short paragraphs; admonitions for callouts; tables for anything enumerable (fields, defaults, thresholds); mermaid for flow/sequence, not for anything a table would do better.
Admonition syntax is mkdocs Material's !!! block form, not GitHub's > [!NOTE] — the two are not interchangeable and only one renders correctly depending on which renderer processes the file (GitHub's own Markdown preview understands > [!NOTE]/> [!WARNING]/> [!CAUTION]; mkdocs-material understands !!! note). docs/docs/** files use the mkdocs form exclusively (verified: grep -rn '^!!! ' docs/docs/features/*.md matches 12 admonitions across 5 files, zero > [! GitHub-style blocks in docs/docs/). CHANGELOG.md and GitHub Release bodies use the GitHub form instead (verified: CHANGELOG.md:163 uses > [!WARNING], shifted from :159 after PR #495 added the deprecation banner), because those render on github.com, not through mkdocs.
Exemplars, quoted directly:
Admonition with a title (docs/docs/features/services.md:110-111):
!!! warning "This is destructive"
All learned priors, correlations, intervals, and aggregates for the selected area are permanently deleted. The integration will start re-learning from scratch on the next analysis cycle (hourly by default). Other areas are unaffected.
Table for a service's return payload (docs/docs/features/services.md, purge_area_history return table):
| Key | Description |
|-----|-------------|
| `area_id` | The area_id that was purged |
| `entities_deleted` | Number of entity rows removed from the database for this area |
```mermaid
sequenceDiagram
participant HA as Home Assistant
participant Coord as Coordinator
...
HA->>Coord: async_config_entry_first_refresh()
Six files under `docs/docs/` use mermaid as of 2026-07-06 (`grep -rl '```mermaid' docs/docs/`): `time-prior-flow.md`, `wasp-in-box.md`, `global-prior-flow.md`, `data-flow.md`, `analysis-chain.md`, `transition-learning.md`. Use `sequenceDiagram` for request/response or component interaction; `flowchart TD` for pipeline/decision steps (see `docs/docs/technical/data-flow.md:40` for a flowchart example).
Avoid hedge-words like "likely" or "may vary — check your instance" in feature docs (an older page, docs/docs/features/decay.md, has this smell in its Output section: "likely as a percentage... Note: The exact representation might vary; check the sensor's state in your HA instance."). If you don't know what a sensor outputs, read the code or run the integration — don't ship a hedge. Newer pages (adjacent-areas.md, sleep-presence.md, services.md) state behavior definitively; hold new/edited pages to that bar and consider it a low-priority cleanup opportunity (not urgent) if you touch decay.md for another reason.
README structure and feature-list conventions
README.md (175 lines, verified 2026-07-06) is structured: badges → hook paragraph → ## The Quick Answer (HA-vs-AOD comparison table) → ## Creating Automations with AOD (workflow + what-it-provides) → ## Documentation (link out) → ## Features → ## Planned Features → ## Installation (HACS + manual) → ## Entities Created → ## Debugging → ## Support & Feedback.
Feature-list convention (## Features, line 75 on): each bullet is **Bold Name**: one-sentence description of user-visible behavior, no implementation detail, no config key names. Sub-bullets are used only for one entry (**Multiple Sensor Support**) that itself enumerates sensor categories. ## Planned Features uses the identical bullet format but for unshipped ideas — before adding something there, check it isn't already shipped (cross-check against ## Features and the actual InputType enum), and before removing something, check it hasn't quietly shipped without the README being updated (the reverse of the #491 problem).
When a feature ships: move its bullet from ## Planned Features to ## Features in the same PR that ships the code, or as an immediate same-day follow-up — don't leave a shipped feature listed as "planned" (this exact gap is why PR #494, "fix broken purpose link in README", exists as a class of fix — the README is a doc-of-record that drifts from code just like everything else).
Docstring convention
Google-style docstrings, full type annotations (Python 3.13+), as directed by CLAUDE.md. Verified against the actual codebase (custom_components/area_occupancy/data/prior.py, utils.py):
defclamp_probability(
value: float, min_val: float | None = None, max_val: float | None = None) -> float:
"""Clamp probability value to valid range.
Args:
value: Probability value to clamp
min_val: Minimum value (default: MIN_PROBABILITY from const)
max_val: Maximum value (default: MAX_PROBABILITY from const)
"""
Module-level docstrings are one-liners or a short paragraph (prior.py's module docstring: """Area baseline prior (P(room occupied) *before* current evidence). followed by a two-sentence elaboration). Not every function needs an Args/Returns block — trivial one-liners (format_percentage, format_value) get a single-line docstring only. Add the full Args:/Returns: block once a function has more than one parameter or a non-obvious return, and always for anything touching the Bayesian calculation (per CLAUDE.md's "100% coverage for calculation changes" instruction — docs and tests should match that rigor).
Commit / PR conventions
Conventional-commit prefixes are used and enforced by convention (not a checked-in linter rule as of 2026-07-06 — verified no commitlint config in repo root): feat, fix, docs, test, refactor, style, chore, optionally scoped, e.g. feat(health):, fix(prior):, docs(adjacent-areas):. Verified from git log --oneline -30:
3471e7a feat(health): purpose-aware stuck-active thresholds and saner defaults (#474)
b9df513 fix(health): preserve user-ignored repairs across condition flaps (#473)
fd61713 docs(adjacent-areas): Phase 5 documentation
8840a50 refactor(analysis): hoist step helpers to module level for C901
Squash-merge titles append the PR number in parentheses, e.g. chore: bump version to 2026.5.17 (#475). Same convention held through the 2026-07-06 merge wave, e.g. feat: adjacent-areas — learned next-door room influence (#454) (verified: current main HEAD as of this sweep, 17b71d2). When writing a PR title, don't add the (#N) yourself — GitHub's squash-merge UI appends it automatically from the PR number.
Release notes and CHANGELOG.md — resolved 2026-07-06, pointer banner in place
Release notes are written by hand in GitHub Releases, not generated from CHANGELOG.md. Verified via gh release view 2026.5.17: a hand-written release body with a narrative intro ("Saner repair defaults + the off-switch you asked for"), a numbered "What's fixed" section explaining each change's motivation, and a "What's Changed" list of feat(scope): ... by @Hankanman in <PR URL> lines matching the actual merged PRs (#472, #473, #474, #475). This is materially richer than a changelog entry — it explains why, tells users what to check (e.g. "Make sure your bedroom area's purpose is set to Sleeping"), and is the artifact users actually read on the releases page and via HACS.
CHANGELOG.md was abandoned (last dated entry ## [2026.3.3] - 2026-03-09, CHANGELOG.md:11, while the actual latest release was 2026.5.17 with several releases in between not represented) — this is now SETTLED. PR #495 (merged 2026-07-06) added a banner immediately under the # Changelog heading: "this file is no longer maintained (last entry 2026.3.3). The changelog of record is GitHub Releases..." This is the lighter-weight version of the two options this doc previously proposed (pointer banner, not a full delete) — don't re-propose deletion or backfilling as if the ambiguity were still open. Do not add new dated entries to the file's body; if a task asks you to log a release there, point to the banner and use GitHub Releases instead.
STALE-DOC landmines — do not trust without re-verifying
These are documents that describe (or described) a process the repo no longer follows. Two of the three rows below were fixed in the 2026-07-06 merge wave (PR #495) — kept here as settled history rather than deleted, since the failure mode (a doc silently diverging from repo reality) recurs and the fix references are useful precedent. Re-check before relying on any row, since things may have drifted again since this sweep.
Doc
What it said
Status as of 2026-07-06
Re-verify with
CONTRIBUTING.md:16
"Fork the repo and create your branch from dev."
SETTLED (PR #495, merged 2026-07-06): now reads "Fork the repo and create your branch from main." No dev branch exists on the remote; all recent feature/fix branches (fix/global-prior-quiet-tail, feat/adjacent-areas, fix/bedroom-half-life-override, etc.) targeted main directly, and the doc now matches.
git branch -r (look for absence of dev/preview); sed -n '16p' CONTRIBUTING.md
CLAUDE.md "Branch and Release Strategy" section
"Development happens on dev branch. PR from dev to preview for prereleases. PR from preview to main for full releases."
STILL A LIVE LANDMINE — CLAUDE.md is unchanged and still says dev→preview→main. The repo has no open PRs as of this sweep (the 2026-07-06 wave, #454/#491/#492/#493/#494/#495/#496, all merged direct to main), and .github/workflows/release.yml triggers on release: published (a manually-created GitHub Release), not a branch-merge event. There is no preview branch either.
git branch -r; cat .github/workflows/release.yml; gh pr list --json baseRefName
pyproject.toml (line 114, now line 113)
Inline comment fail_under = 85 # Enforce 90% coverage minimum — the enforced number and the comment's stated number disagreed with each other in the same line.
SETTLED (PR #495, merged 2026-07-06): now reads fail_under = 85 # Enforced global minimum; aim for 90%+ on core calculation modules (CLAUDE.md) — internally consistent and cross-references CLAUDE.md's "85%+ coverage requirement (90% for core calculations)" line instead of contradicting it.
grep -n fail_under pyproject.toml
Do not silently "fix" CLAUDE.md's branch-strategy section as a drive-by edit — it's explicit project instruction content; flag it to the maintainer or fold the correction into a PR whose primary purpose is a docs/process cleanup, per aod-change-control's rules on touching CLAUDE.md. (The other two rows show this is a real, fixable pattern once someone owns it — CLAUDE.md's branch-strategy section is the one that hasn't been picked up yet.)
Keeping docs in sync when code changes — the #491 cautionary example
Cautionary example (PR #491, fix(prior): keep quiet tail in global prior denominator, merged 2026-07-06 — verify with gh pr view 491): the code change altered how PriorAnalyzer.calculate_and_update_prior() picks the end of its observation period (previously truncated to last_interval_end when the area had been quiet >1h; now always now — this is the current, permanent behavior on main, i.e. actual_period_end is always now, not conditionally truncated). docs/docs/technical/global-prior-flow.md documented the old (buggy) behavior in prose — "If last interval is more than 1 hour old: Use last_interval_end / Otherwise: Use current time" — and had to be edited in the same PR (7 additions, 16 deletions) to stop describing the bug as intended behavior. This was caught as a review nitpick, not by the original author remembering to update the doc.
Rule extracted: when a change touches data/analysis.py, data/prior.py, data/decay.py, or utils.py::bayesian_probability() (the files CLAUDE.md names under "Modifying Bayesian Calculation"), grep docs/docs/technical/ and docs/docs/features/ for any prose description of the specific mechanism you changed before opening the PR:
If a technical page describes the old behavior, update it in the same PR — don't wait for a reviewer to catch it. This applies most to technical/global-prior-flow.md, technical/time-prior-flow.md, technical/bayesian-calculation.md, technical/calculation-flow.md, technical/decay content inside deep-dive.md, and technical/transition-learning.md.
When a change needs docs at all
Use this checklist before deciding a PR doesn't need a doc touch:
Does it change a formula, threshold, default, or constant a technical page states as fact? → update the relevant technical/*.md.
Does it change what a user configures, sees, or can do (new sensor type, new service, new config field, new sensor entity)? → update the relevant features/*.md and, if user-visible enough, the README ## Features bullet list.
Does it ship something previously listed under README ## Planned Features? → move the bullet.
Does it change a config-flow step, migration behavior, or anything CONTRIBUTING.md/CLAUDE.md describes procedurally? → flag for a separate docs/process PR rather than silently drifting further (see landmines above).
Is it a pure internal refactor with no behavior change (e.g. refactor(analysis): hoist step helpers to module level for C901)? → no doc update needed; say so explicitly in the PR description so a reviewer doesn't go looking for one.
Did you add a new feature? → write the features/ + technical/ pair together (see adjacent-areas.md / transition-learning.md above), and add both to mkdocs.yml's nav:.
Provenance and maintenance
Date-stamped 2026-07-06 (post-merge sweep), main HEAD 17b71d2 (feat: adjacent-areas — learned next-door room influence (#454)). Integration release version is still 2026.5.17 — the 2026-07-06 merge wave (#454, #486, #488, #489, #491–#496) is on main but not yet in a tagged release; don't describe any of that wave's changes as "shipped in release" until a new tag/release exists.
Re-verification commands by volatile fact category:
Docs tree / nav completeness: find docs/docs -name '*.md' | sort vs grep -oE '[a-z-]+/[a-z-]+\.md' docs/mkdocs.yml | sort -u
