finding-skills

Finding skills

Use this skill when the user is looking for a capability that might exist as an installable skill — "how do I work with PDFs here", "find a skill for code review", "add the pdf skill from anthropics/skills" — or when you wish you had domain instructions you don't have.

Where skills come from

1. agentstack search <query> — the user's central library, the embedded catalog, and the official MCP Registry, in one pass. Results marked (in manifest) are already installed.
2. Any skills repo on GitHub/GitLab — the ecosystem publishes SKILL.md directories, and agentstack add skill speaks its conventions:

agentstack add skill anthropics/skills --list        # inspect a repo's skills
agentstack add skill anthropics/skills --skill pdf   # preview one (dry run)
agentstack add skill owner/repo@pdf                  # same, alias spelling
agentstack add skill https://github.com/o/r/tree/main/skills/pdf
agentstack add skill ./local-skill
agentstack more lib add owner/repo --skill pdf            # into the central library

Everything previews first. The dry run fetches into transient staging, scans the content, and shows the manifest diff plus the exact digest that would be pinned — nothing persistent changes until a human re-runs with --write.

Judge quality by evidence, not popularity

There are no install counts here, on purpose. The signals that matter:

• Scan verdict — the preview shows per-skill findings. High-severity findings (hidden Unicode) block the add; warnings deserve a read before you recommend proceeding.
• Pin + provenance — after a write, the lockfile pins the exact commit and content checksum, and agentstack more explain <name> shows where a skill came from and whether its content still matches its pin.
• Description quality — a skill without a frontmatter description is invisible to search and to agents; treat that as a smell.

Rules for agents

• Propose, don't apply: run the dry run, show the user the preview, and let them re-run with --write. Never pass --allow-flagged yourself — a blocked scan finding is the user's decision.
• Prefer the central library for skills the user will want across repos (agentstack more lib add …, then reference by name in profiles); prefer the project manifest for repo-specific skills.
• After a write in a gateway-served (zero-files) project, remind the user that trust re-gates on the edit: they run agentstack trust . themselves — never run it for them.
• When nothing suitable exists, say so and offer to help directly; a new skill directory with a SKILL.md (name + description frontmatter) is all it takes to make the solution reusable.