> ## Documentation Index
> Fetch the complete documentation index at: https://humboldt.metaphi.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Skills

> A directory with a `SKILL.md`. Humboldt loads one when the task matches.

## Location

| Path                     | Scope                           |
| ------------------------ | ------------------------------- |
| `.agents/skills/<name>/` | Repository                      |
| `.claude/skills/<name>/` | Repository (Claude Code layout) |
| `~/.hum/skills/<name>/`  | You                             |

## Evaluate and improve

```bash theme={"dark"}
hum skill eval <name>                    # run every case in a fresh workspace
hum skill propose <name> --action append_section --target "Nothing to do" --content note.md --rationale "..." --evidence <run-dir> <run-dir>
hum skill apply <name> <proposal>        # re-run cases; revert if any pass -> fail
hum skill standing <name>                # outcomes per task class, eval history
```

## Cases

One JSON per case under `<skill>/cases/`:

| Field       | Meaning                               |
| ----------- | ------------------------------------- |
| `task`      | The instruction                       |
| `gate`      | Command run afterwards; exit 0 passes |
| `fixtures`  | Directory copied in first (optional)  |
| `max_turns` | Ceiling (optional)                    |

Keep one case where the right answer is to do nothing. A proposal needs two runs as evidence.