skill-scout

Scout public agent-skill directories (ranked by installs) and propose new skills for the-grid, deduped against what is wired. Use when the user says "skill-scout", "find new skills" or "what skills sh

SKILL.md

GitHub

skill-scout

Find popular Claude skills the-grid doesn't have yet, and propose them — ranked by install count. Propose only. Never add submodules or run wire.sh; the user adds skills via the-grid's normal flow (git submodule add … && bash scripts/wire.sh).

How to run

  1. Gather candidates.

    • Ranked web directories: WebFetch each listing page and pull the ranked rows — skill name, popularity number (installs), and source owner/repo. Normalise install counts ("1.8M" → 1_800_000) so they sort.
    • Local library submodules: read them straight off disk — categories/*.md for awesome-list indexes, or grep SKILL.md frontmatter for real collections. These have no install count — keep them in a separate, unranked pile.
  2. Load what the-grid already has. From the repo root (GRID_DIR, default the repo this skill lives in):

    • Root-level skills: top-level directories containing a SKILL.md.
    • Submodule skills: repos/*/**/SKILL.md (one or more skill dirs per submodule).
    • Read each skill's name: from its frontmatter. Also note the source repo of each submodule (git config -f .gitmodules or the repos/<name> path) so a match on either skill name or source repo counts as "already have it".
  3. Dedupe. Drop any candidate whose skill name OR source repo the-grid already carries. Matching is case-insensitive; compare on the bare skill name (ignore owner/ prefixes) and on the owner/repo source.

  4. Rank and propose. Sort the ranked-directory survivors by install count, descending. Present the top ~15 as a table:

    Skill Installs Source (owner/repo) Why it fits the-grid

    For "Why it fits", write one honest line covering fit for a developer workflow (dev workflow, debugging, git, planning, testing, infra/ops). If a candidate is a poor fit, say so plainly rather than inventing a reason — or omit it.

    Then, separately, surface a handful of relevant picks from the local index (repos/voltagent) grouped by category — no install column, since it isn't ranked. Flag that these are OpenClaw-ecosystem skills worth a manual look, not ranked recommendations.

  5. Hand off, don't act. End with the exact commands the user would run to add any they pick, but do not run them:

    git submodule add <repo-url> repos/<name>
    git submodule update --init
    bash scripts/wire.sh
    

Notes

  • "Based on installs" means the directory's own popularity ranking — we don't recompute it. If a source has no popularity signal, it doesn't belong here.
  • Skills already in the-grid via a submodule (e.g. repos/anthropic/skills/*, repos/gstack/*) must be deduped by source repo too — a popular skill from vercel-labs/skills is "new" only if that repo isn't already a submodule.
  • Keep the proposal short and skimmable. The user decides; this skill only scouts.

Sections

Read a section only when its condition applies; do not preload them.

  • sections/sources.md: read when gathering candidates (step 1)