capture-lesson

Files (1)

• SKILL.md

SKILL.md (first lines)

# Capture lesson (harness memory)

Turn a mistake into a short, reusable rule. This is how the harness "gets better" without model fine-tuning.

## Steps

1. Infer from the conversation (or user args):
   - **Symptom** — what went wrong
   - **Root cause** — why
   - **Rule** — what to do differently next time (imperative, 1–3 sentences)
   - **Scope** — language/repo/tool if specific
   - **Anti-pattern** — what not to repeat

2. Append one entry to:

`$env:USERPROFILE\.grok\lessons\LESSONS.md`

Use this template:

```markdown
## YYYY-MM-DD — <short title>

- **Symptom:** …
- **Cause:** …
- **Rule:** …
- **Scope:** …
- **Anti-pattern:** …
```

2b. Regenerate the always-loaded digest:

`python "$env:USERPROFILE\.grok\skills\skill-router\scripts\lessons-digest.py"`

It rewrites `~/.grok/rules/03-lessons.md` with the top 12 rules. Do not edit that generated file by hand.

3. If the lesson is project-specific and the user is in a git repo, also offer to add a one-liner under that repo’s `AGENTS.md` (only if they agree).

4. If the mem0 MCP tools are available, store a one-line version as a memory (add_memory). Otherwise suggest `/remember` if Grok memory is enabled.

5. Confirm the file path and the rule in one sentence. Do not lecture.

## Quality bar

- Actionable for a future agent with no chat history
- No secrets, tokens, or personal data
- Prefer concrete ("always run X before Y") over vague ("be careful")

From TheNonFungibleHuman/multi-model-agent-harness, MIT.

More from TheNonFungibleHuman/multi-model-agent-harness