agentleFS
the-Nile-jonesthe-Nile-jones/claude-skills/skills/skill-scout/SKILL.md

skill-scout

Skill0 starsChanged 4 months agoMITView on GitHub

Recon-and-triage loop for external skill/pattern sources. Use when Nile wants to scan a skill hub, repo, marketplace, or article for new skills / skill ideas / atomic candidates / diminishing-returns — "scan for new skills", "anything new in anthropics/skills", "check the skill hubs", "mine this source for skill ideas", "what should I grab", "/skill-scout". Fetches the source live, Venn-dedups every candidate against installed skills AND the atom library FIRST, then emits three buckets: new-skill candidates, atomic candidates, already-covered (diminishing returns). It does NOT install, build, or mint atoms — those are separate gated steps it hands off to. The dedup gate is the load-bearing part; it exists because improvised scans keep skipping it and "discovering" things already owned. On-demand: suggest it, don't auto-fire.

23 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use context-budget.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • get-proofpilotSkill

    context-budget

    Audits Claude Code context window consumption across agents, skills, MCP servers, and rules. Identifies bloat, redundant components, and produces prioritized token-savings recommendations.

    Also in 3 projects that use this

The file

We split this skill into the rules it gives an agent

2 rules. 0 are common, 0 are distinctive, and 2 are very specific to this project.

  • 0

    Common

    Most of the 22 edited versions we read keep these.

  • 0

    Distinctive

    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
  1. Very specific:

    Install: when it's a packaged Anthropic/community skill or an MCP that exists and audits clean → /mcp-install (MCPs) or note the install …

    line 53
  2. Very specific:

    Build: when it's a *pattern* with no clean package, or it needs adaptation to Nile's stack → skill-creator.

    line 54

“Common” is measured: kept by at least 6 in 10 of the 22 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

skill-scout — External Source Recon & Triage

A deterministic loop for "is there anything here worth taking?" — applied to skill repos, plugin marketplaces, cookbooks, or a single article. It exists because the same procedure kept getting hand-rolled, and the hand-rolled version skips the dedup gate and reports things you already own as new. (Lived example: a teaching-loop "gap" that was already atomized from the same page that day.) The gate is the whole point.

What it produces — always these three buckets

  1. New-skill candidates — capabilities NOT covered by an installed skill. Per candidate: one line of what it does, the closest existing skill, gap type (none / partial / full), and an install-vs-build call.
  2. Atomic candidates — structural principles worth porting to the Exocortex (not whole skills). These are flagged, not minted — they are the input to /pattern-extraction, which does the actual RMA/DMA + atom write. skill-scout never writes to the atom store itself.
  3. Already-covered / diminishing-returns — what the source restates that you already have, each named to its covering skill or atom. This bucket is the deliverable, not the leftover — it's the signal that tells you a vein is mined so you stop spending energy on it.

The loop

1. Identify, classify, and ground

Three things before fetching anything substantive:

  • Owned or community? Anthropic-owned sources (github.com/anthropics/skills, anthropics/anthropic-cookbook, the Claude Docs skills pages, the /claude-api bundled list) carry a low fabrication-prior. Everything else — community marketplaces, awesome-* aggregators, obra/superpowers, third-party blogs — is Tier-4: apply source-vetting, strip fabricated stats and capability claims before crediting anything. Note the owner explicitly; it sets the skepticism level. Caveat: "trusted" lowers the prior, it does not exempt falsifiable claims (API behavior, version specifics) from verification against a primary tool like /claude-api. The prefill-400 claim was true but still had to be checked.
  • Already audited? Check for a prior audit of this exact source — vault (KB/Claude/Best Practices/*audit*), the MA validation log, or a previous skill-scout breadcrumb. If scanned before, run in DIFF MODE: report only what changed since the last audit. Don't trust the prior audit blindly, though — spot-check that a few items it marked "covered" are still installed (an audit can outlive the reality it described).
  • Load the fit-context — non-negotiable, and the easiest step to skip. Every "no-fit" / "install" / "build" call you'll make depends on knowing Nile's stack and what's been evaluated-and-removed before. A cold session that skips this will mis-triage (e.g. flag pptx as a candidate when slides have no workflow here). Load it from: ~/.claude/rules/skill-router.md (the removed-skills + demoted-skills history — this is the canonical "already judged" record), ~/CLAUDE.md (stack + canonical tools), and ls ~/.claude/skills/. A candidate that matches a removed skill goes straight to bucket 3 with the removal reason, never bucket 1.

2. Fetch the source live and enumerate candidates — never from memory

Pull the current state (WebFetch, or gh api for GitHub trees — training-data recall is stale and violates verify-before-asserting). Enumerate per source-type:

  • Repo / hub → list the actual skill directories (gh api repos/<o>/<r>/git/trees/<branch>?recursive=1). GitHub's HTML file-tree is JS-rendered and won't enumerate via WebFetch — use the API.
  • Article / blog → extract every concrete technique or prompt verbatim; there are no directories, the "candidates" are the techniques.
  • Marketplace → enumerate listings; raise the vetting bar (Tier-4 by default, often unmaintained).

3. Pre-filter large candidate sets — and log what you cut

For a handful of candidates, skip to step 4. For a large set (a marketplace, a sprawling repo), the per-candidate gate doesn't scale — so cut obvious no-fits by category first (creative-coding, slides, enterprise team-comms, anything matching a removed skill from step 1). Then state what you dropped and why — Pre-filtered N candidates by category: [list]. Silent truncation reads as "scanned everything" when it didn't; an explicit drop-list keeps the cut honest and lets Nile overrule a category.

4. Targeted dedup gate — the load-bearing step

For each surviving candidate, query the baseline for that candidate's domain (don't pre-pull the whole library — targeted beats snapshot):

  • Does an installed skill already do this? (you have the ls from step 1; read the matching skill's description.) → bucket 3 (name it), or bucket 1 gap=partial only if the candidate adds a real delta.
  • Is the underlying principle already an atom? → mcp__neo4j-cypher__read_neo4j_cypher, matching family / tags / statement keywords for this candidate's domain. "Skill ideas" most often turn out to be principles you already hold. → bucket 3 (name the atom), or hand the delta only to /pattern-extraction.
  • Absent from both? → bucket 1 (skill) or bucket 2 (atomic candidate).

Skipping this for even one candidate is how false "new" findings get through. If something "obviously" looks new — that's exactly when to run it.

Rigor floor — ≥3×3× per atomic scan (standing rule, 2026-06-15). The gap/diagonal verdict for an atomic scan target is not a glance — run it through at least a 3×3× RMA/DMA pass (compose /pattern-extraction): 3 constructive passes decomposing what the candidate actually offers, 3 adversarial passes attacking the claim (already-covered? off-stack? survives Test-11 grounding in Nile's actual regime, not the abstract? a real delta or just a restatement of something owned?). Upgrade up to 10×10× when the juice is worth the squeeze — a high-value target (an official vendor skill for a daily-use in-stack tool, a plausible diagonal) earns the deeper pass; a niche/personal-tool catch-all does not. Calibrate the squeeze to the prize. (The 2026-06-15 Cloudflare eval ran ~1×1× and got lucky; this floor exists so the next verdict doesn't depend on luck.)

5. Triage + install-vs-build (bucket 1 only)

  • Install when it's a packaged Anthropic/community skill or an MCP that exists and audits clean → /mcp-install (MCPs) or note the install path. Community installs get /skill-security-auditor first.
  • Build when it's a pattern with no clean package, or it needs adaptation to Nile's stack → skill-creator.

6. Report + leave a dedup breadcrumb

Emit the three buckets (template below). Then append a one-line audit entry — source, date, verdict, candidate counts, and the drop-list from step 3 — where the next scan will check in step 1 (the capture buffer for boot-sweep, or a KB/Claude/Best Practices/*-audit-*.md doc for a substantial source). A scan that doesn't record itself dooms the next scan to re-mine.

Report template

# skill-scout — <source> (<owned|Tier-4>) — <date> [<full scan | diff since <prior audit>>]

## New-skill candidates (N)
- <name> — <what it does>. Closest existing: <skill or "none">. Gap: <none|partial|full>. → <install | build>

## Atomic candidates (N)  → hand off to /pattern-extraction
- <principle, one line>. Closest atom: <id or "none">. Delta worth porting: <what>

## Already-covered / diminishing returns (N)
- <source item> = <covering skill/atom id>   (or: = REMOVED <skill>, <reason>)

## Pre-filtered (if any)
- <N candidates dropped by category: ...>

## Verdict
<one line: is this vein worth more passes, or mined?>

Boundaries — what skill-scout does NOT do

  • Doesn't install, build, or mint. It triages and reports. Installing (/mcp-install, skill-creator) and atom-writing (/pattern-extraction) are separate gated actions it hands off to — keeping recon read-only means a scan can never have a surprising side effect.
  • Division of labor with /pattern-extraction: scout is triage — "is there anything here worth mining?" — and is read-only. pattern-extraction is extraction — it runs the full RMA/DMA and writes atoms. Scout's bucket-2 atomic candidates are pattern-extraction's input, not a duplicate of its Step 0. Run scout to decide whether a source is worth a pattern-extraction pass; run pattern-extraction to actually mine it. Don't run both as redundant dedups on the same source.
  • Doesn't trust Tier-4 at face value, or trusted sources blindly. Community sources get per-claim vetting; Anthropic-owned sources still get their falsifiable claims checked against a primary tool.
  • Doesn't re-mine. Already-audited source → diff mode, with a spot-check that the prior audit still holds.

Provenance

Captures the scan-and-triage procedure run ad hoc 3× in one session (NotebookLM docs, XDA engineer prompts, skill-hubs) before being made deterministic — the skill-candidate rule (a repeatable procedure seen 3+ times → route it to skill-creator rather than filing it as another one-off task). Dedup discipline mirrors a duplicate-source registry and the source-vetting rule above. Hardened via a 5×5 RMA/DMA pass 2026-06-15 (fixes: cold-session fit-context loading, fetch-before-baseline ordering, large-source pre-filter with logged drops, pattern-extraction boundary, per-source-type modes, trusted≠unverified).

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.