Codex Prompt Batching

Purpose

This playbook explains how Codex, Cursor, terminal agents, or future local AI agents should triage a large prompt or a batch of prompts before editing files.

Use it when one human message contains several independent requests, touches many ownership areas, or could be split across parallel branches.

Batch Triage Rules

Split the prompt into independent tasks.
Estimate touched paths for each task.
Mark conflict risk.
Assign parallelizable tasks to different branches or worktrees.
Run non-parallelizable tasks serially.
Serialize core protocol, root config, app registry, and design token changes.
Give every task acceptance criteria.
After each task, output changed files, commands run, risks, and follow-ups.
If a prompt is too broad, create a task plan before editing the whole repository.

Prompt Batch Triage Table

| Task | Scope | Paths | Conflict Risk   | Parallelizable | Branch | Merge Order |
| ---- | ----- | ----- | --------------- | -------------- | ------ | ----------- |
| 1    |       |       | low/medium/high | yes/no         |        |             |

Conflict Risk Guide

Risk	Examples	Execution
Low	Docs-only updates, isolated app copy, one Storybook story, leaf mock data.	Can run in parallel with a scoped branch.
Medium	Shared UI additions, app package changes, mock API skeletons, app registry display-only edits.	Parallel only with disjoint paths and a merge plan.
High	App contract, window manager, shell routing, app registry composition, design tokens, root config, database, deployment.	Serialize and require human review.

Acceptance Criteria Template

## Acceptance Criteria

- Scope stays within:
- User-visible behavior:
- Docs updated:
- Tests/checks:
- Not included:

Worker Output Template

## Result

- Branch:
- Changed files:
- Commands run:
- Architecture impact:
- Docs impact:
- Risk:
- Follow-ups:

Relationship To Task Graphs

Normal Direct Implementation Mode does not require a task graph. Use this batching playbook for lightweight triage.

Use docs/development/ai-control-plane/task-graphs/ only when the human explicitly asks for opt-in multi-agent orchestration, slicing, RFC-style planning, or staged execution.