Grok CLI Template

# AGENTS.md — Project Instructions

Grok CLI auto-loads this file from the project root before any work. Keep it small and durable — push deep workflows into Skills under `.agents/skills/`.

## Meta-Learning Protocol

You operate under The Learner Brain methodology. Improve the *system*, not just the output.

### Upgrade Triggers
1. **Repetition** — same correction twice
2. **Friction** — workaround the user shouldn't repeat
3. **Manual labor** — multi-step task that should become a Skill
4. **Preference** — codify expressed style/approach

### Upgrade Format
```
🧠 BRAIN UPGRADE PROPOSAL
Trigger: [what caused this]
Surface: [AGENTS.md | Skill | knowledge/<category>]
Location: [path]
Proposed change: [exact addition]
Rationale: [why it prevents future friction]
```

### Hygiene
- Keep AGENTS.md under ~200 lines
- Never edit AGENTS.md without approval — propose ONE upgrade at a time
- Prefer Skills over inline procedures for anything > 10 lines

## Project Context

### Tech Stack
- Language: [Your language]
- Framework: [Your framework]
- Test command: [bun test | pnpm test | etc.]

### Working Agreements
- Run lint + tests before claiming a task is done
- Ask before adding production dependencies
- Never commit `.env*` or anything under `secrets/`
- Use sub-agents (`task`/`delegate`) for research-heavy work to keep main context clean

## Patterns Learned
<!-- Append discovered patterns. Promote to Skills when they grow. -->
- [Pattern]: [solution]

## Anti-Patterns
- Do not regenerate boilerplate when an existing Skill covers it
- Do not run `computer_*` tools without explicit approval

---
name: meta-learning
description: Detect repetition, friction, manual labor, and unwritten preferences. Propose ONE Brain Upgrade at a time using the Learner Brain format.
---

# Meta-Learning Skill

Triggers (any of):
1. The user corrects the same thing twice in this session
2. The user describes a workaround they shouldn't have to repeat
3. A task takes more than 3 manual steps that could be codified
4. The user states a preference that isn't written down anywhere

When a trigger fires, STOP and emit:

```
🧠 BRAIN UPGRADE PROPOSAL
Trigger: <what caused this>
Surface: <AGENTS.md | Skill | knowledge/<category>>
Location: <path>
Proposed change: <exact text to add>
Rationale: <why it prevents future friction>
```

Wait for explicit approval ("add it", "do it") before editing any file.

## Outputs
- One proposal per turn, max
- Reference existing entries when relevant — don't duplicate

## Version
- 1.0.0 — Initial

// ~/.grok/user-settings.json — Grok CLI user settings
// Configure default model, sub-agents, and MCP servers
{
  "defaultModel": "grok-code-fast-1",
  "subAgents": {
    "explore": {
      "description": "Read-only repo exploration; returns a concise summary",
      "model": "grok-4-1-fast-reasoning"
    }
  },
  "mcpServers": {
    // Add MCP server configs here, e.g.:
    // "context7": { "command": "npx", "args": ["-y", "@upstash/context7-mcp"] }
  }
}