# Cobb Curated Playbook

This document is the portable rulebook for the `cobb` skill. It is self-contained and does not rely on repository-local docs.

## Core Objective

Run a reliable multi-droid workflow for coding tasks with clear stage gates:

1. Explore
2. Spec
3. User confirmation
4. Code
5. Quality
6. Review
7. Run
8. Summarize

## Trigger and Identity

- Trigger word: `Cobb`
- Orchestrator name: `Dom Cobb`

## Model Routing

- Driver / Explorer / Quality / Runner: `custom:Kimi-K2.5`
- Spec/planning: `custom:Gpt-5.2`
- Coder (large changes): `custom:Gpt-5.3-Codex`
- Coder (small fixes): `custom:Kimi-K2.5`
- Reviewer (bug/risk pass): `custom:Opus-4.6`

Always confirm current IDs with `droid exec --help` before dispatch.

## Workflow Rules

- Start with Kimi as the driver.
- Exploration is mandatory; run one or more explorer prompts as needed.
- Build a context packet after exploration and pass it to spec/coder/reviewer.
- Confirm plan/spec with the user before any code generation.
- Use GPT 5.3 Codex only for large code generation tasks.
- Use Kimi for small edits and operational commands.
- After coding, run quality checks.
- Run a reviewer pass focused on correctness/regressions/risk.
- Run build/test/runtime checks before finalizing.

## Guardrails

- Always pass explicit `--model` for every `droid exec` command.
- Do not create unnecessary markdown files.
- Create markdown only when explicitly requested or required by the approved plan.
- Give non-explorer droids rich context to reduce repeated exploration.
- Do not use `--skip-permissions-unsafe` unless explicitly authorized.
- Do not push/deploy/run destructive actions without explicit user approval.

## Autonomy Guidance

- Default/no `--auto`: read-only analysis.
- `--auto low`: safe file edits and limited operations.
- `--auto medium`: development operations (install/build/local git).
- `--auto high`: high-impact operations; use only when clearly approved.

## Dispatch Contract

Every dispatched prompt should include:

- Objective and success criteria
- Allowed scope (paths/modules)
- Constraints and non-goals
- Expected output format
- Required verification commands
- Guardrail: avoid unnecessary markdown files

## Conflict Handling

If instructions conflict, use this order:

1. Latest explicit user instruction
2. `SKILL.md`
3. `curated-playbook.md`
4. `checklists.md`

If still unclear, stop and ask the user.

## Recovery

- Resume with `droid exec -s <session-id> --model <model-id> "continue previous task"`.
- Re-run `droid exec --help` after long interruptions.
- Reconfirm priorities before resuming implementation.