Audit the list
An Audit re-runs the admission test over live sections of README.md: every entry re-verified against current data, tiers reassessed, structure reconsidered, then the maintainer adjudicates and the result is committed. The rules live in CONTRIBUTING.md (Admission, Evidence, Entry Ordering) and CONTEXT.md (vocabulary) — this skill is the process that applies them, not a second copy of them. Entry changes land only on the maintainer's explicit go.
1. Scope
Resolve the scope from the arguments. Named sections mean exactly those, whether or not they were audited before (a re-run is how a past verdict gets rechecked). all is ambiguous — AskUserQuestion whether it means every section or only never-audited ones; prior audits are recorded in git history (git log --oneline --grep="sweep\|audit"). With no arguments, AskUserQuestion listing the never-audited Thematic Groups. Batch the work one Thematic Group per sitting. Done when the section list is settled.
2. Evidence
Fetch live evidence for every entry in scope before judging anything (CLAUDE.md verification rule):
- Downloads/month:
cd website && uv run python fetch_pypi_downloads_via_clickpy.py — free keyless ClickPy sweep of the full README, sole writer of data/pypi_downloads.tsv (rewritten from scratch each run; header row, every row stamped with its fetched_at date). A cache whose fetched_at is within the last 7 days is current enough for verdicts — skip the sweep; older than that, re-run it (costs ~1s). Cross-checks print to stdout, take explicit names, and never touch the cache: fetch_pypi_downloads_via_bigquery.py <name> ... (canonical source, maintainer's own GCP account, --dry-run first — the docstring carries the cost constraints; full-README sweeps exceed the free tier, keep name lists small), fetch_pypi_downloads_via_pepy.py <name> ... (needs PEPY_TECH_API_KEY in repo-root .env, throttled to 5 requests/minute), or https://pypistats.org/api/packages/{name}/recent paced 8s or slower. pypistats excludes mirror/CI traffic; ClickPy, BigQuery, and pepy include it — never mix sources within one comparison.
- Repo state: archived flag, last push, created date, stars, description —
gh api repos/{owner}/{repo}, GitLab API for GitLab-hosted projects.
- PyPI metadata (
https://pypi.org/pypi/{name}/json) wherever a name might not be the canonical package — ownership collisions and wrong display names surface here. Wrong-package cache rows are common: the sweep looks up README display names, so when the display name differs from the canonical package, the row silently measures an unrelated squatter or a dead predecessor. When a famous entry shows a count that looks off (too small, or absurdly round), verify identity before citing it, and fetch the canonical package's count via pepy.
- Renaming an entry can silently expose it to the sweep, and no automatic check catches that. The sweep only queries names matching
PYPI_NAME_RE, so a display name carrying a space or other non-PyPI shape is skipped outright. Rename it into a PyPI-shaped one and it starts being queried: uv audit became uv-audit, and PyPI uv-audit is an unrelated third-party tool by rocshers, not Astral's subcommand. After any rename, fetch https://pypi.org/pypi/{new-name}/json and confirm the package is the project the entry links to; when it is not, add a null entry to website/data/pypi_name_overrides.json with the reason. Only one case is guarded automatically: test_bundled_entries_are_never_queried_on_pypi requires a null override for every (part of X) entry. Two broader checks were measured against the full list and rejected — do not re-propose either. Requiring a package's PyPI metadata to link back to the entry's GitHub repo misses this case completely, since uv-audit declares no repo URL at all, as do 26 legitimate entries including numba, selenium and pyglet; it also flags 10 benign org moves and binding splits (gensim, msgspec, pendulum, instructor, duckdb). Flagging entries whose display name differs from their repo name returns 46 hits, every one legitimate, and uv-build sits among them with the shape identical to uv-audit despite being a real Astral package.
Done when every entry in scope has downloads (or a stated no-signal reason), repo state, and a confirmed PyPI name.
3. Verdicts
Draft a verdict and reason for every entry, restructure before cap: decide Splits, mints, and re-homes first — cutting before restructuring destroys entries — then apply the cap tier by tier, including promotions and demotions between obvious choice and challenger. Ground every reason in the fetched evidence; label anything unverifiable as a judgment call. Where fresh evidence contradicts the standing verdict, say so in the reason instead of silently keeping the seed.
4. Review and go
Run the preview-verdicts skill: it generates the interactive review page and defines how the maintainer's feedback JSON comes back and gets processed. Their verdicts are final. Before touching README.md, surface what the feedback implies but does not decide — cap overflows, homeless entries after a Split, tier for a tierless flip — as named assumptions or questions, then get an explicit go.
5. Execute
One commit per section: body lists each removal with its reason and downloads figure; restructures, tier moves, and reorders ride the same commit. Format-only outcomes (no removals) are a single style commit. make test before every commit, make build after the last one. Generic commit helpers tend to split a section audit into structural and per-subcategory commits — if that happens, squash back to one commit per section. Done when the tree is clean, tests passed before each commit, and the build count reconciles with the adjudicated changes.
6. Record
A conclusion that outlives the sitting goes into the repo before the audit ends: admission or evidence rules into CONTRIBUTING.md, repo process and environment facts into CLAUDE.md and AGENTS.md (kept in sync), vocabulary into CONTEXT.md. Done when nothing the next audit needs is stranded in conversation.
1---2name: audit-the-list3description: Audit README.md sections against the shortlist rules — re-verify every entry's verdict with live data, promote or demote challengers, restructure oversized use cases, prune, and evaluate proposed additions. Fire when the maintainer asks to audit, sweep, prune, re-check, or maintain sections, or asks whether an entry (or a proposed one) deserves its slot.4---5
6# Audit the list
7
8An Audit re-runs the admission test over live sections of README.md: every entry re-verified against current data, tiers reassessed, structure reconsidered, then the maintainer adjudicates and the result is committed. The rules live in CONTRIBUTING.md (Admission, Evidence, Entry Ordering) and CONTEXT.md (vocabulary) — this skill is the process that applies them, not a second copy of them. Entry changes land only on the maintainer's explicit go.
9
10## 1. Scope
11
12Resolve the scope from the arguments. Named sections mean exactly those, whether or not they were audited before (a re-run is how a past verdict gets rechecked). `all` is ambiguous — `AskUserQuestion` whether it means every section or only never-audited ones; prior audits are recorded in git history (`git log --oneline --grep="sweep\|audit"`). With no arguments, AskUserQuestion listing the never-audited Thematic Groups. Batch the work one Thematic Group per sitting. Done when the section list is settled.
13
14## 2. Evidence
15
16Fetch live evidence for every entry in scope before judging anything (CLAUDE.md verification rule):
17
18- **Downloads/month**: `cd website && uv run python fetch_pypi_downloads_via_clickpy.py` — free keyless ClickPy sweep of the full README, sole writer of `data/pypi_downloads.tsv` (rewritten from scratch each run; header row, every row stamped with its `fetched_at` date). A cache whose `fetched_at` is within the last 7 days is current enough for verdicts — skip the sweep; older than that, re-run it (costs ~1s). Cross-checks print to stdout, take explicit names, and never touch the cache: `fetch_pypi_downloads_via_bigquery.py <name> ...` (canonical source, maintainer's own GCP account, `--dry-run` first — the docstring carries the cost constraints; full-README sweeps exceed the free tier, keep name lists small), `fetch_pypi_downloads_via_pepy.py <name> ...` (needs `PEPY_TECH_API_KEY` in repo-root `.env`, throttled to 5 requests/minute), or `https://pypistats.org/api/packages/{name}/recent` paced 8s or slower. pypistats excludes mirror/CI traffic; ClickPy, BigQuery, and pepy include it — never mix sources within one comparison.
19- **Repo state**: archived flag, last push, created date, stars, description — `gh api repos/{owner}/{repo}`, GitLab API for GitLab-hosted projects.
20- **PyPI metadata** (`https://pypi.org/pypi/{name}/json`) wherever a name might not be the canonical package — ownership collisions and wrong display names surface here. Wrong-package cache rows are common: the sweep looks up README display names, so when the display name differs from the canonical package, the row silently measures an unrelated squatter or a dead predecessor. When a famous entry shows a count that looks off (too small, or absurdly round), verify identity before citing it, and fetch the canonical package's count via pepy.
21- **Renaming an entry can silently expose it to the sweep, and no automatic check catches that.** The sweep only queries names matching `PYPI_NAME_RE`, so a display name carrying a space or other non-PyPI shape is skipped outright. Rename it into a PyPI-shaped one and it starts being queried: `uv audit` became `uv-audit`, and PyPI `uv-audit` is an unrelated third-party tool by rocshers, not Astral's subcommand. After any rename, fetch `https://pypi.org/pypi/{new-name}/json` and confirm the package is the project the entry links to; when it is not, add a null entry to `website/data/pypi_name_overrides.json` with the reason. Only one case is guarded automatically: `test_bundled_entries_are_never_queried_on_pypi` requires a null override for every `(part of X)` entry. Two broader checks were measured against the full list and rejected — do not re-propose either. Requiring a package's PyPI metadata to link back to the entry's GitHub repo misses this case completely, since `uv-audit` declares no repo URL at all, as do 26 legitimate entries including numba, selenium and pyglet; it also flags 10 benign org moves and binding splits (gensim, msgspec, pendulum, instructor, duckdb). Flagging entries whose display name differs from their repo name returns 46 hits, every one legitimate, and `uv-build` sits among them with the shape identical to `uv-audit` despite being a real Astral package.
22
23Done when every entry in scope has downloads (or a stated no-signal reason), repo state, and a confirmed PyPI name.
24
25## 3. Verdicts
26
27Draft a verdict and reason for every entry, restructure before cap: decide Splits, mints, and re-homes first — cutting before restructuring destroys entries — then apply the cap tier by tier, including promotions and demotions between obvious choice and challenger. Ground every reason in the fetched evidence; label anything unverifiable as a judgment call. Where fresh evidence contradicts the standing verdict, say so in the reason instead of silently keeping the seed.
28
29## 4. Review and go
30
31Run the `preview-verdicts` skill: it generates the interactive review page and defines how the maintainer's feedback JSON comes back and gets processed. Their verdicts are final. Before touching README.md, surface what the feedback implies but does not decide — cap overflows, homeless entries after a Split, tier for a tierless flip — as named assumptions or questions, then get an explicit go.
32
33## 5. Execute
34
35One commit per section: body lists each removal with its reason and downloads figure; restructures, tier moves, and reorders ride the same commit. Format-only outcomes (no removals) are a single style commit. `make test` before every commit, `make build` after the last one. Generic commit helpers tend to split a section audit into structural and per-subcategory commits — if that happens, squash back to one commit per section. Done when the tree is clean, tests passed before each commit, and the build count reconciles with the adjudicated changes.
36
37## 6. Record
38
39A conclusion that outlives the sitting goes into the repo before the audit ends: admission or evidence rules into CONTRIBUTING.md, repo process and environment facts into CLAUDE.md and AGENTS.md (kept in sync), vocabulary into CONTEXT.md. Done when nothing the next audit needs is stranded in conversation.
vinta