Agent directives

Boundaries

Respect the architecture's seams. Examples:

• Desktop shells: main/renderer talk over IPC only.
• Servers: routes → controllers → services → data. No skipping layers.
• UI frameworks: no side effects outside effects/lifecycle hooks.

1. Reuse before adding

• Search for an existing utility before writing one.
• Priority: standard library > existing dependency > new dependency (see package-choice).

2. Read before grep

1. Read {{DOCS_DIR}} and the folder's AGENTS.md / CLAUDE.md first.
2. Then targeted searches with file globs.
3. Never blind-scan the whole tree.

3. Plan, then build

• Plan → approve → implement → verify → commit.
• Tests first where it pays: failing test → implement → green.

4. Verify before done

• Run the module's check after every change.
• Typecheck / static analysis must be clean.
• A check you could not run is reported as not run, never as passing.

5. Never

• Put secrets, tokens or customer data in code, docs or logs.
• Edit generated files by hand; regenerate them.
• Leave the work order's scope without saying so.

Customize for your project

Module	Stack	Verify command
{{MODULE}}	{{STACK}}	{{VERIFY_COMMAND}}

Add your hard out-of-scope list (things the product deliberately never does).