Discover and load skills¶

Find the skill that covers a task, load it for your agent, and author, evaluate, share and improve skills of your own.

Prerequisites¶

A signed-in CLI.
For authoring: a coding agent installed, for evaluations.

Find and load a skill¶

Search by meaning, not keywords. Quote the query:
```
❯ alis skills search "deploy a service to production" --json --limit 3
```
The result's queriedSkills[] carries id, displayName, description, version, source and samplePrompt.
Check the fit cheaply before spending tokens on the whole document:
```
❯ alis skills preview <id>
```
Load it. The markdown prints verbatim; an agent reads it and follows it over guesswork:
```
❯ alis skills load <id> --json
```
JSON
```
{"markdown": "…"}
```
Inside a coding session the load is attributed to the session automatically ($CLAUDE_SESSION_ID, then $ALIS_SESSION_ID). Offline, a previously loaded copy is served from the cache with a note on stderr.
Fetch a file the instructions reference, with the path exactly as written:
```
❯ alis skills resource <id> references/checklist.md
```
alis skills resource <id> --all --out <folder> writes the whole bundle.

Nothing fits? Ask the Alis Build team for one:

❯ alis skills request --name "…" --description "…" --use-case "…"

alis skills list is for browsing by eye; alis skills explain describes the whole system and, with --json, lists every capability with its command.

Keep a local copy for your agent¶

Skills load live, so nothing needs installing. When you want a complete copy in your agent's native skill folder:

Shell

alis skills install <id> --harness claude
alis skills installed
alis skills upgrade --all
alis skills uninstall <id>

Author a skill¶

Create a workspace. It lives under ~/.alis/skills/workspaces/<id>/:
```
❯ alis skills create my-skill --description "…" --sample-prompt "…"
```
Or start from work you just finished: alis skills capture my-skill --session last. Or import a public GitHub folder: alis skills import https://github.com/<owner>/<repo>/tree/<branch>/<folder>.
Edit SKILL.md (and optional references/, scripts/, assets/). Put the phrases people would type in double quotes in the description.
Evaluate it with real headless agent runs:
Shell
```
alis skills evaluate my-skill --init
alis skills evaluate my-skill --baseline --json
```
Results land in evals/runs/<timestamp>/; evals/ never publishes.
Publish:
```
❯ alis skills publish my-skill
```
Publish validates the front matter, blocks on detected secrets and refuses the starter's TODO sample prompt. Exit 3 with remote_changed means the registry moved since you installed the workspace: --force overwrites, or alis skills edit my-skill --refresh rebases.

Share it. New skills start private:

Shell

alis skills share my-skill --domain
alis skills share my-skill --email ana@example.com

Buildable code (templates, a Dockerfile, Terraform) belongs in a code block, not in a skill; the block's .skill/ folder becomes its skill. A skill that does carry templates/ is materialised with alis skills apply <id> --into <folder> --set name=value --yes.

Give feedback and improve a skill¶

Report a gap. It opens a conversation with the skill's maintainers:
```
❯ alis skills feedback <id> --rating down "the deploy step skips --plan-only"
```
alis specialist send-session --ticket <t> attaches the agent session that hit it.
As the owner, turn the feedback into a revision with your agent, then evaluate and publish:
Shell
```
alis skills improve <id> --ticket <t> --json
alis skills evaluate <id> --baseline
alis skills publish <id>
```
Publish drafts a note on each answered feedback conversation. Not the owner? alis skills improve <id> --propose forks it and alis skills propose <id> sends the owner your patch.
Read the trail:
```
❯ alis skills history <id>
```

In the TUI, ShiftL opens the Skills page and + starts the guided create, edit, evaluate, publish, share walkthrough.