skill-find git:20260724.13fed82A

Immutable. This exact content never changes and is served at /api/v1/blob/beb4c541a3f454a7.

---
name: skill-find
description: Discover, try without installing, and install agent skills — first from the official Fastfold catalog (github.com/fastfold-ai/skills), then from the open skills.sh ecosystem (npx skills find / use). Use when the user wants to find, search, browse, list, try/use without installing, add, enable, disable, or remove skills/capabilities (e.g. fold, MD, protein design, Google Calendar).
---

# Skill Find

> **This skill has no scripts.** Do not run `python scripts/...` for it. Use the
> `skills.manage` tool (actions: `find`, `list`, `info`, `install`, `remove`,
> `enable`, `disable`), and only fall back to the `sandwalk skills ...` CLI or
> `npx skills ...` when the tool is unavailable. Ephemeral try-without-install
> uses `npx skills use` (there is no `skills.manage` action for that yet).

Help the user discover agent skills, try them once without installing, or install
the ones they want. Search in two tiers — the trusted Fastfold catalog first, then
the broader open ecosystem — and always scan live, because both catalogs grow over time.

- **Fastfold catalog (first-party, trusted):** https://github.com/fastfold-ai/skills (skills under `skills/<name>/`)
- **Open ecosystem (skills.sh):** https://skills.sh — the package manager for community skills ([vercel-labs/skills](https://github.com/vercel-labs/skills))

## When to Use This Skill
- The user asks to "find skills", "search skills", "what skills are available", "list sandwalk skills", or "add a skill".
- The user asks to "try a skill", "use a skill without installing", or "run this skill once".
- The user asks "how do I do X", "is there a skill for X", or "can you do X" where X is a specialized capability the agent lacks.
- The user wants to enable, disable, or remove an installed skill.
- The user wants to extend the agent with folding, MD, protein design, reporting, or other packaged workflows.

## Workflow

### Tier 1 — Fastfold catalog (try this first)
Scan live with the agent tool (catalog + ecosystem when `query` is set):

```
skills.manage(action="find")                   # Fastfold catalog listing
skills.manage(action="find", query="protein")  # catalog + public ecosystem (npx)
skills.manage(action="find", query="google calendar")
```

Present `name`, `install_source`, and `description` as a short numbered list.
Never auto-pick an ecosystem skill — confirm with the user first.

### Tier 2 — Suggested provider collections
When the Fastfold catalog lacks a capability, also consider these curated third-party
collections:

- **K-Dense-AI** — `K-Dense-AI/scientific-agent-skills`
- **Anthropic** — `anthropics/life-sciences`
- **DeepMind** — `google-deepmind/science-skills`

### Tier 3 — Open ecosystem (skills.sh)
`skills.manage(action="find", query="...")` already runs `npx -y skills find <query>`
when Node is available. If npx is missing, tell the user to install Node or browse
https://skills.sh. Prefer official/audited sources and higher install counts.

Do **not** use the desktop Skills UI Vercel login path from the agent — that catalog
browser stays Integrations → Vercel only.

### Confirm
Ask whether to **try once** (no install) or **install** for reuse. Ecosystem skills
run third-party code; call that out before using or installing anything outside the
Fastfold catalog. You do **not** need to ask the user to confirm again in chat as a
gate — `skills.manage` install/remove pauses for desktop UI or interactive CLI
approval automatically. `npx skills use` does not persist files into Sandwalk skill
dirs (ephemeral temp prompt only).

### Use without installing (ephemeral)
Prefer this when the user wants a one-off try, or has not decided to keep the skill.
`npx skills use` resolves the same sources as `npx skills add`, writes skill files to
a temp directory, and prints a generated prompt to **stdout**. Capture that output
(via shell) and follow it for the current task — do **not** pass `--agent` (that
starts an external coding agent interactively).

```bash
# owner/repo@skill-name → prompt on stdout
npx -y skills use <owner>/<repo>@<skill-name>

# equivalent with explicit --skill
npx -y skills use <owner>/<repo> --skill <skill-name>

# Fastfold catalog example
npx -y skills use fastfold-ai/skills@fold
# or
npx -y skills use fastfold-ai/skills --skill fold
```

After the prompt is printed, execute those instructions in this session (read any
paths the prompt references under the temp skill dir if needed). Tell the user the
skill was **not** installed; offer `skills.manage(action="install", ...)` if they
want it permanently.

### Install
```
skills.manage(action="install", source="fastfold-ai/skills@skills/<name>")
skills.manage(action="install", source="<owner>/<repo>@<subpath>")
skills.manage(action="install", source="https://github.com/<owner>/<repo>/tree/<ref>/<subpath>")
skills.manage(action="install", source="/path/to/skill-dir")
skills.manage(action="install", source="/path/to/skill.zip")   # or .skill bundle
```

If the tool returns `blocked: true`, ask the user to run `sandwalk skills add <source>`
or `sandwalk config set skills.allow_agent_install true`.

### Enable / disable / remove
```
skills.manage(action="list")
skills.manage(action="disable", name="<skill-name>")
skills.manage(action="enable", name="<skill-name>")
skills.manage(action="remove", name="<skill-name>")
```

### Confirm availability
Installed or re-enabled skills are picked up on the **next user message** (the system
prompt reloads each turn) — not mid-same-turn. They also appear in the composer `/`
menu after the UI refreshes. Verify with `skills.manage(action="list")`.

**Always include the skill slash mention in your reply** (e.g. `/fold` or
`/lab-notes`) so the desktop UI renders a clickable chip to skill details — same
as when the user inserts a skill via `/` in the composer.

## Sources You Can Install From
- Fastfold catalog (whole repo): `fastfold-ai/skills`
- Fastfold catalog (one skill): `fastfold-ai/skills@skills/<name>`
- Any GitHub URL: `https://github.com/<owner>/<repo>/tree/<ref>/<subpath>`
- GitHub shorthand: `<owner>/<repo>@<subpath>`
- A local path to a skill directory
- A local `.zip` or `.skill` archive containing `SKILL.md`

## Notes
- Installs use native `git clone` and fall back to `npx skills add` when needed.
- Try-without-install uses `npx -y skills use <source>` (see
  [vercel-labs/skills](https://github.com/vercel-labs/skills)); install still goes
  through `skills.manage` / `sandwalk skills add`.
- Both catalogs are dynamic — re-run discovery instead of assuming a fixed set.
- Default to the Fastfold catalog; reach for the open ecosystem only when needed.