| name | find-skills |
| description | Discover external skills during GedPi's skill-fit checkpoint. When the current bundled, project, and user skills don't cover a task's needs, search the open ecosystem via npx skills, evaluate candidates, and install through GedPi's project-skill mechanism. Triggers include "how do I do X", "find a skill for X", "is there a skill for X", "skill for", "need a skill", "search for skills", "skill-fit", "ecosystem" |
Find Skills: Ecosystem Discovery for GedPi's Skill-Fit Checkpoint
Part of GedPi's skill-fit workflow. In subagent-enabled workflows, ged-explorer normally performs the read-only inventory/evaluation/search and reports findings to the main brain; the main brain then decides whether to install or create project skills. In solo/no-subagent workflows, the main brain performs the whole checkpoint itself.
Use after grill-me (if needed) and before skill-creator. The flow is:
clarify → explorer read-only inventory/search (or main brain in solo mode) → main-agent install/create decisions → plan
Resist the urge to skip straight to skill-creator when an off-the-shelf skill exists. Conversely, don't exhaustively search when the gap is narrow and project-specific — go straight to skill-creator.
When to Use
During the skill-fit checkpoint (non-trivial tasks)
When the inventory of bundled/project/user skills doesn't cover the clarified task, search the ecosystem for an existing skill before creating one from scratch.
When the user asks directly
Use this when the user asks for help with a specific domain or capability — anything from "how do I write tests for this" to "I need a skill for React patterns."
How GedPi Discovers Skills
GedPi loads skills from three tiers, in this priority order (last wins on name conflicts):
| Tier | Location | Source |
|---|
| Bundled | skills/<name>/SKILL.md | Shipped with GedPi |
| Project | .ged/project-skills/<name>/SKILL.md | Managed per-project via GedPi workflow |
| User | ~/.agents/skills/<name>/SKILL.md | npx skills add -g or manual |
Use npx skills list to see what's globally available, and check .ged/SKILLS.md for what's currently installed/recommended in the project.
The Skill-Fit Checkpoint
During non-trivial task planning, before drafting the plan:
- Clarify the task enough to judge skill needs.
- Inventory and select what's available — bundled, project-scoped, and user skills. With subagents enabled,
ged-explorer should do this read-only reconnaissance.
- Search with this skill if coverage has gaps — search only for the missing capability. With subagents enabled,
ged-explorer may run npx skills find and report candidates, but must not install anything.
- Decide/install/create in the main brain: install project-scoped external skills if warranted, or create with
skill-creator if no adequate skill exists and the gap is reusable at the project level.
- Proceed to planning.
Never install global/user skills automatically. The GedPi workflow installs skills into .ged/project-skills/ (project-scoped, cleaned up when tasks are done). External installation requires user consent.
How to Search the Ecosystem
1. Check the leaderboard first
Browse skills.sh to see top-ranked skills. High install counts (>1K) signal battle-tested packages.
2. Run the CLI search
npx skills find [query]
Queries should focus on the missing capability, not the general domain:
| Too broad | Better |
|---|
npx skills find react | npx skills find react performance |
npx skills find testing | npx skills find playwright e2e |
npx skills find python | npx skills find python testing |
3. Evaluate each result
| Check | What to look for |
|---|
| Install count | ≥1K is solid, <100 is risky |
| Source reputation | Vercel Labs, established OSS authors |
| GitHub stars on source repo | <100 stars → skepticism |
| Content quality | Does the SKILL.md describe concrete patterns or vague advice? |
Only flag a skill as viable if its SKILL.md content meaningfully fills the gap. A skill that's tangentially related but doesn't actually help with the task at hand is noise.
How to Install
There are three paths depending on context:
A. Project-scoped install (GedPi workflow)
When the skill fits a task in the current plan, GedPi's ensureTaskSkillDependencies() copies the skill's content into .ged/project-skills/<name>/SKILL.md and tracks it in SKILLS-STATE.json. The skill is automatically cleaned up when no open task depends on it.
This is the primary path — it keeps skills scoped to the project that needs them.
B. User-global install (explicit user consent only)
npx skills add <owner/repo>@<skill> -g -y
Installs to ~/.agents/skills/<name>/ where Pi discovers it across all projects. Only use this when:
- The user explicitly asks to install globally
- The skill is general-purpose (e.g., language patterns, testing frameworks)
- You've presented the option and the user approved
Do not auto-install globally during the Ged workflow.
C. Project-local creation (no adequate external skill)
When the search turns up nothing suitable and the gap is reusable within the project, use skill-creator to author a narrow project-local skill. This keeps the knowledge scoped to .ged/project-skills/ without polluting the global namespace.
Common Search Categories
| Category | Example queries |
|---|
| Web Development | react nextjs, typescript css, tailwind design |
| Testing | playwright e2e, vitest, jest mocking |
| DevOps | docker compose, ci-cd github actions |
| Documentation | api docs, changelog, readme |
| Code Quality | code review, lint, refactoring |
| Design | ui ux, accessibility, design-system |
| Productivity | git workflow, automation, project setup |
Presenting Results to the User
When you find relevant skills, present them clearly:
Found a skill that fits: "react-performance" from vercel-labs/agent-skills
- Install count: 185K
- Covers: React rendering optimization, memo patterns, bundle analysis
- Source: github.com/vercel-labs/agent-skills
Two install options:
1. Project-scoped (recommended): I'll set it up in .ged/project-skills/
2. Global (npx): npx skills add vercel-labs/agent-skills@react-performance -g
Let the user decide. If they choose project-scoped, GedPi handles the install.
When No Skills Are Found
If the search returns nothing useful:
- Acknowledge the gap clearly
- Assess whether it's a one-off task (just do it with general capabilities) or a reusable pattern
- If reusable, offer to create a project-local skill via
skill-creator
I searched for "rust embedded no-std patterns" and didn't find a match.
This looks like a one-off task — I can handle it directly.
If it comes up again, I can create a project-local skill for it.
Guiding principles
- Search narrow, not broad — query for the missing capability, not the whole domain
- Prefer project-scoped installs — keeps skills tied to the projects that need them
- Don't auto-install globally — user consent required for
npx skills add -g
skill-creator is the fallback — when no external skill fits, create a project-local one- YAGNI for skills too — don't install skills for hypothetical future needs
