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

# Claude Code Skills

> Three skills that teach Claude Code how to bootstrap, audit, and explain your evals — installed with one command.

multivon-eval ships three [Anthropic Agent Skills](https://docs.claude.com/en/docs/agents-and-tools/agent-skills) so Claude Code knows when to bootstrap an eval suite, when to gate a PR on it, and how to explain why a particular evaluator was picked. The skills are plain Markdown with YAML frontmatter (no DSL), and they live in `multivon_eval/_skills/` inside the PyPI package.

## What is an Agent Skill?

A skill is a directory containing a `SKILL.md` file that Claude Code auto-loads from `~/.claude/skills/`. The frontmatter declares `name`, `description`, and `allowed-tools`; the body is instructions Claude Code reads when the description matches the user's request. Without skills, Claude Code has to infer evaluator selection and CLI flags from your docs, and hallucinates command names half the time. With skills, the tool's own team writes the workflow once and every Claude Code session inherits it.

## The bootstrap to audit to explain loop

The three skills form one iterative loop. You start cold, run bootstrap, ask explain to surface what just got picked, then on every subsequent PR audit gates the change against the suite bootstrap generated.

<CardGroup cols={3}>
  <Card title="eval-bootstrap" icon="rocket" href="/skills/eval-bootstrap">
    Cold-start an eval suite from a product description plus sample traces. Emits `eval_suite.py`, `seed_cases.jsonl`, `thresholds.yaml`, `DISCOVERY_REPORT.md`.
  </Card>

  <Card title="eval-audit" icon="shield-check" href="/skills/eval-audit">
    Pre-ship gate on a PR diff. Runs only the cases that stress the changed surface. Blocks safety-class regressions at p \< 0.05.
  </Card>

  <Card title="eval-explain" icon="message-circle-question" href="/skills/eval-explain">
    Three-sentence answer to "why did multivon pick this evaluator". Reads `DISCOVERY_REPORT.md` plus the evaluator docstring.
  </Card>
</CardGroup>

## Install

The skills ship inside the `multivon-eval` PyPI package (>= 0.9.8). One command writes them into `~/.claude/skills/`.

```bash theme={null}
pip install 'multivon-eval>=0.9.8'
multivon-eval install-skills
```

<Note>
  `install-skills` prefers a directory symlink into `~/.claude/skills/eval-bootstrap` (and the two siblings) so a later `pip install -U multivon-eval` propagates `SKILL.md` edits without re-running the command. On Windows or filesystems that refuse directory symlinks, it falls back to a recursive copy and prints a note that you'll need to re-run the command after upgrades.
</Note>

Two flags worth knowing:

<ParamField path="--dry-run" type="flag">
  Print what would happen — which source paths, which targets, symlink vs copy — without touching the filesystem. Run this first if you're unsure what `~/.claude/skills/` already contains.
</ParamField>

<ParamField path="--force" type="flag">
  Replace existing entries at the three target paths. Without `--force`, the command refuses to overwrite anything already on disk and tells you which entries collided.
</ParamField>

## Auto-discovery flow

Once the symlinks exist under `~/.claude/skills/`, Claude Code auto-loads them on session start. No config edit is needed; a new session picks them up on next launch. Each session's tool list includes the skill names; Claude Code matches user phrases against each skill's `description` and invokes the matching skill before answering. The auto-invoke triggers are spelled out on each skill's detail page.

Verify the install:

```bash theme={null}
ls ~/.claude/skills/
# Expect: eval-audit  eval-bootstrap  eval-explain  (plus any others you have)
```

## Manual fallback

If you can't run `install-skills` (older versions, or you want to vendor into a different directory), the symlinks are three lines:

```bash theme={null}
PKG_PATH=$(python -c "import multivon_eval, pathlib; print(pathlib.Path(multivon_eval.__file__).parent)")
mkdir -p ~/.claude/skills
# rm first: `ln -sf` against an existing symlinked DIRECTORY creates a
# nested link inside the target instead of replacing it.
rm -rf ~/.claude/skills/eval-bootstrap ~/.claude/skills/eval-audit ~/.claude/skills/eval-explain
ln -s "$PKG_PATH/_skills/eval-bootstrap" ~/.claude/skills/eval-bootstrap
ln -s "$PKG_PATH/_skills/eval-audit"     ~/.claude/skills/eval-audit
ln -s "$PKG_PATH/_skills/eval-explain"   ~/.claude/skills/eval-explain
```

## What you can do once installed

Three concrete prompts that route to the three skills:

<CodeGroup>
  ```text Cold-start a project theme={null}
  "Add evals to this project."

  → Routes to eval-bootstrap. Reads pyproject.toml to detect your
    LLM provider, finds sample traces (traces/, data/traces/, or
    prompts you), runs `multivon-eval bootstrap`, rewrites
    `stub_model` to call your real model, writes EVALS.md.
  ```

  ```text Gate a PR theme={null}
  "Will this prompt change regress evals?"

  → Routes to eval-audit. Scopes the diff via
    `multivon_eval.attribution.scan`, identifies stressed cases,
    runs them with `python eval_suite.py --runs 3`, reports
    PASS / WARN / BLOCK with Wilson CI + paired-McNemar.
  ```

  ```text Understand a pick theme={null}
  "Why did multivon recommend Faithfulness here?"

  → Routes to eval-explain. Reads DISCOVERY_REPORT.md, the
    evaluator's docstring, and an example seed case. Answers
    in exactly 3 sentences plus an illustrative case.
  ```
</CodeGroup>

The three together close the loop from cold project to gated CI without you ever leaving Claude Code.