AGENT FUEL MAP FOR AN AI-NATIVE IDE

Use this before giving an agent access to a project.

1. Name the task

[ ] Chore: formatting, documentation, a tiny test, or isolated cleanup.
[ ] Bug: a repeatable failure with a log or failing test.
[ ] Refactor: behavior must stay stable and the affected area is named.
[ ] Product decision: interface, copy, architecture, public API, or customer effect.
[ ] High risk: authentication, secrets, money, deletion, production, or licensing.

2. Choose the model route

[ ] Use a cheaper or local model for bounded chores, first-pass search, and repeatable changes.
[ ] Use a stronger model for unclear debugging, architecture, security review, and product decisions.
[ ] Require a person for money, secrets, account actions, production writes, legal decisions, merge, deploy, send, or delete.

3. Set the boundary

Working directory:
Files and folders allowed:
Tools allowed:
Network destinations allowed:
Forbidden areas:

4. Set the brakes

[ ] Read-only work may run without review.
[ ] Review file writes before they land.
[ ] Review installs and network calls before they run.
[ ] Run terminal and fetch work in a sandbox.
[ ] Require a person before merge, deploy, send, or delete.

5. Require a receipt

[ ] Change summary.
[ ] Tests, logs, or screenshots.
[ ] Sources used.
[ ] Model and tools used.
[ ] Cost and retry note.
[ ] Rollback path.
[ ] Open questions and unresolved risks.

Want this applied to your setup? The $49 PimpMyIDE audit turns screenshots and one repeated workflow into a one-page diagnosis with three ranked fixes.
https://pimpmyide.com/#audit