Skill Search

Use this helper when an agent needs one of two discovery strategies:

1. Skill search — accept a query and return only the top-k keyword matches. Keeps the full catalog out of context; scales to large catalogs.
2. Standard progressive disclosure — return a compact index of every skill (name, description, path) upfront, then load the full SKILL.md body for whichever skill the agent selects.

Configuration

• my_skills_path: optional user-managed skills root. Repeat if more than one user root is needed.

The helper resolves the central catalog in this order: the --central-root flag, then the SKILL_SEARCH_CENTRAL_ROOT environment variable, then a sibling ../skills directory. To make this skill expose the Genesis catalog, symlink skill-search/ into your agent's skills dir and (if it isn't beside skills/) point one of those at the catalog. The alternative — flattening every skill into a client skills directory for direct discovery — is handled by the repo's unpack.sh (./unpack.sh --help).

Skill search

Use when the agent should search on demand rather than hold the full catalog in context:

python scripts/skill_search.py --query "draft a Slurm batch script for a four-GPU job"

With a user-provided skills path and a custom result limit:

python scripts/skill_search.py \
  --query "validate Croissant metadata for an ML dataset" \
  --top-k 3 \
  --my-skills-path /path/to/my/skills

Scoring weights name token matches (×2) over description token matches (×1), with substring bonuses for closer matches. Results with a score below --min-score (default 1) are excluded.

Standard progressive disclosure

Use when the agent should receive the full menu upfront and pick without needing a query. The compact index (names + descriptions + paths, no bodies) is loaded first; the full SKILL.md body is loaded on demand for the selected skill.

python scripts/skill_search.py --load-all

With a user-provided skills path:

python scripts/skill_search.py --load-all --my-skills-path /path/to/my/skills

After loading the index, load a selected skill's body via the Python API:

from skill_search import load_skill_body
body = load_skill_body(selected_skill)

Pass --include-prompt to also receive a pre-formatted <available_skills> XML block suitable for injection into an agent system prompt.

Output

The script returns machine-readable JSON so it can be called directly by agents or host tools. Unreadable catalog entries appear in the skipped array and as warnings on stderr; stdout remains valid JSON.

Python API

If the agent runtime can import Python packages directly, the skill_search package provides:

• search_payload(query, top_k, min_score, ...) — skill search: returns top-k scored matches
• progressive_disclosure_payload(...) — standard progressive disclosure: returns compact index of all skills
• load_skill_body(skill) — load the full SKILL.md body for a selected skill
• discover_skills(paths) — discover all skills from one or more roots
• to_prompt(skills) — generate an <available_skills> XML block
• build_search_callable() — create a reusable search function with pre-configured roots

See README.md for details.

Related Files

• scripts/skill_search.py
• README.md