Agents must work in the smallest useful scope. These limits apply to every session unless the user explicitly widens scope.
File and change limits
LOW tasks: touch at most 3 files; prefer 1–2
MEDIUM tasks: touch at most 7 files; list them in the plan before editing
HIGH tasks: do not edit product code until the user approves a written plan
Line budget
Prefer patches under 150 changed lines per iteration
If a change exceeds 300 lines, split into smaller PRs or stop and propose a sequence
When to STOP
Stop after analysis (no edits) for:
Auth, billing, tenant isolation, or database schema changes
Cross-cutting refactors without an approved file list
Tasks where the agent would need to scan most of the repository
Enforcement
Use rg, fd, and targeted reads — never load entire directories into context
If scope grows during work, pause and ask before continuing
Raw content
Copy into your project — e.g. .instructions.md, .agent.md, or SKILL.md
## Overview
Agents must work in the smallest useful scope. These limits apply to every session unless the user explicitly widens scope.
## File and change limits
- **LOW tasks:** touch at most 3 files; prefer 1–2
- **MEDIUM tasks:** touch at most 7 files; list them in the plan before editing
- **HIGH tasks:** do not edit product code until the user approves a written plan
## Line budget
- Prefer patches under 150 changed lines per iteration
- If a change exceeds 300 lines, split into smaller PRs or stop and propose a sequence
## When to STOP
Stop after analysis (no edits) for:
- Auth, billing, tenant isolation, or database schema changes
- Cross-cutting refactors without an approved file list
- Tasks where the agent would need to scan most of the repository
## Enforcement
- Use `rg`, `fd`, and targeted reads — never load entire directories into context
- If scope grows during work, pause and ask before continuing