curate-cluster
Skill49 starsChanged 13 days agoApache-2.0View on GitHub
Reconcile a research or documentation cluster's index with files on disk, repair dangling references, refresh counts and context, and ignore regenerable artifacts. Commit only quiescent documentation, experiment, tool, and index paths; exclude trees peers are actively changing. Use after a documentation burst, when an index drifts, or on a recurring gardening cadence.
In this repository151 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- .github/copilot-instructions.mdCopilot instructions
- llms-full.txtllms.txt
- llms.txtllms.txt
- agent-readinessSkill
- appeal-scoreSkill
- bottleneck-mapSkill
- ci-repairSkill
- claim-repro-scoreSkill
- clear-out-wipSkill
- commit-cleanSkill
- conflation-scoreSkill
- cross-validateSkill
- debt-cleanSkill
- debt-orchestratorSkill
- disambiguation-scoreSkill
- dojo-rsi-scoreSkill
- dos-dispatch-loopSkill
- dos-dispatchSkill
- dos-next-upSkill
- dos-replanSkill
- dos-witness-claimSkill
- field-borrowSkill
- fleet-waveSkill
- git-subagent-syncSkill
- goalSkill
- guard-rsi-scoreSkill
- harness-creatorSkill
- harness-gardenSkill
- industry-scoreSkill
- issue-orchestratorSkill
- issue-queueSkill
- issue-triageSkill
- lightgap-scoreSkill
- memory-compactSkill
- milestone-scoreSkill
- modularizeSkill
- negframe-scoreSkill
- opencode-waveSkill
- operator-heaviness-scoreSkill
- persona-scoreSkill
- phased-planSkill
- plan-auditSkill
- quality-scoreSkill
- question-loopSkill
- refresh-cachedoc-numbersSkill
- refresh-readmeSkill
- releaseSkill
- resume-watchdog-auditSkill
- run-it-all-nightSkill
- score-2xSkill
- scorecardSkill
- scout-loopSkill
- signal-firstSkill
- skill-lifecycleSkill
- skill-overlapSkill
- skill-scoreSkill
- slop-scoreSkill
- sota-checkSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 2 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 25
Markdown mid-edit is self-healing; code mid-edit is not: Before committing, find every file-tree that changed in the last ~60 s.
- Very specific: line 93
Add any new generated kind to .gitignore, not to a commit.
- Distinctive: line 102
If any code lane reports LIVE, exclude that tree from the commit.
- Very specific: line 121
Keep the commit body to a few bullets: index entries added, counts/context fixed, artifacts ignored.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
curate-cluster — the index-and-commit gardening pass
What this does. Keeps the project's index doc true to what is on disk, keeps build artifacts out of git, and commits the curation work — without colliding with any other agents editing this repo at the same time. When the tree is a live multi-session checkout (several sessions building or curating in parallel), a naive
git add -A && git commitwill snapshot a peer's half-written code. This pass is built to never do that: it stages only by explicit path.
The shape: survey disk → reconcile the index → gitignore artifacts → detect live lanes → commit only the quiescent curation lane → verify.
The one rule that matters (learned the hard way)
Markdown mid-edit is self-healing; code mid-edit is not. Before committing, find every file-tree that changed in the last ~60 s. Treat them by kind:
- Code under active construction (any
*.go/*.py/ source tree a peer is building) → exclude it. It is another worker's lane; its intermediate state may not compile and is not yours to commit. Leave it for its builder. - Markdown / data / index (
*.md,experiments/**,tools/**) → safe to commit even if a peer just touched it: a slightly-stale index entry self-heals on the next pass; a mid-write source file does not.
You never need a lock. git's own index serializes the commit; you only need to
choose what to stage — and you stage it by explicit path.
Step 1 — Survey the disk
# top-level docs + scripts that should be indexed
ls *.md *.py 2>/dev/null
# tracked vs untracked
git status --porcelainMap what each kind is: top-level prose lives in *.md; models/scripts in *.py;
runnable experiments under experiments/NN-name/; reusable tooling under
tools/; rendered diagrams under visuals/. The project's index doc names the
sections.
Step 2 — Reconcile the index doc against disk
Three checks (all read-only) against the project's index doc (INDEX.md here):
# (a) docs/scripts on disk but NOT referenced in the index → need an entry
for f in $(ls *.md | grep -v '^INDEX.md$') *.py; do
grep -q "$f" INDEX.md || echo "UNREFERENCED: $f"
done
# (b) filenames referenced in the index that don't exist → dangling (fix/remove)
grep -oE '`[A-Za-z0-9_.-]+\.(md|py)`' INDEX.md | tr -d '`' | sort -u | while read f; do
[ -e "$f" ] || echo "DANGLING: $f"
done
# (c) counts in the header vs reality
echo "docs on disk: $(ls *.md | grep -v '^INDEX.md$' | wc -l) scripts: $(ls *.py | wc -l)"A prose phrase like "a skill.md file" can show as a false-positive DANGLING —
it is text, not a file reference. Ignore those.
For each UNREFERENCED doc, add an entry in the house format (match the neighbours exactly), placed in the right section and cross-linked. Mirror whatever template the index already uses for its entries — title, filename, a short description, the load-bearing result, structure, status, and related docs.
Also refresh, when stale: the Generated date, the Index covers: N documents count line, the Git context blockquote, and any non-document artifacts bullets. Add the new doc to the numbered Reading order if it belongs on a main thread.
Step 3 — Gitignore regenerable artifacts
Compiled / generated outputs must never be committed — they bloat history and
churn. Confirm .gitignore covers them:
git check-ignore experiments/**/bench.exe # should print the path (ignored)Add any new generated kind to .gitignore, not to a commit.
Step 4 — Detect live lanes (the concurrency gate)
# any source tree touched in the last 60s = a peer is building it → exclude
find . -name '*.go' -o -name '*.py' -newermt '-60 seconds' -printf '%p (LIVE)\n' 2>/dev/nullIf any code lane reports LIVE, exclude that tree from the commit. Only commit code you intend to commit when its build is green — never a red or actively-written tree.
Step 5 — Commit the quiescent curation lane (explicit path only)
Stage every doc/index/tool path you curated by explicit path via fak commit --path — never
git add -A, never an exclude-glob sweep. List exactly the files you touched:
fak sync reconcile --apply
# guard: confirm no live code lane or ignored artifact slipped in
git status --porcelain | grep -iE '\.exe$|\.dos/' && echo VIOLATION || echo OK
fak commit \
--path INDEX.md --path <each curated doc> --path <each tool> \
-m "Curate cluster: <one-line of what changed>" [--push]
fak sync pushKeep the commit body to a few bullets: index entries added, counts/context fixed,
artifacts ignored. If you deliberately left an actively-built lane to its builder,
say so in one line. Do not add a Co-Authored-By trailer.
Step 6 — Verify (witness, don't assume)
git log -1 --oneline
git status --short # remaining = only the peer lanes you deliberately leftThe pass is done when: the index has no UNREFERENCED/DANGLING entries, the header counts match disk, no artifact/live-code was committed, and the only remaining working-tree changes are the lanes you intentionally left to their builders.
Running it on a cadence
This is a single pass. To keep the cluster continuously clean, drive it with
/loop (e.g. /loop curate-cluster, self-paced) so it re-surveys after each
burst of doc-writing. Each pass is idempotent — if nothing drifted and nothing
is quiescent-new, it commits nothing and exits.
Where this misleads / honest limits
- The 60 s liveness window is a heuristic, not a lease. For a hard guarantee use
a real file-tree lease (e.g.
dos arbitrate) before committing. - Entry quality is the model's judgment; the format checks are mechanical but "is this the load-bearing number?" is not.
- It assumes the project's house index format. A different cluster needs its own entry template in Step 2.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

