Dialect

Field guide

How each lab wants to be asked.

Distilled from public system-prompt patterns — XML vs markdown, NEVER lists, tool declarations, plan-then-act. Not a dump of anyone's hidden prompt.

File scope

Name the surface. Unscoped agents wander.

When: Any code change in a real repo.
Before

Fix the dashboard.

After

Edit only the dashboard shell and sidebar. Do not touch auth or data layer.

Plan then act

Coding agents have a plan mode for a reason. Use it on anything bigger than a one-liner.

When: Multi-file work, or you don't trust the first instinct.
Before

Rebuild the portal nav and the household view and the ADV export.

After

Plan only. List files, risks, and a sequence. Do not edit until I say go.

Fence the ask

Labs write 'don't add features beyond what was asked' because agents gold-plate.

When: Every implementation prompt.
Before

Add a loading state.

After

Add a loading state to the household table. No new libraries, no extra skeleton variants, no docs.

Verify loop

Tell it the check, or it will declare victory from a compile.

When: UI, data, or anything a user can see.
Before

Make sure it works.

After

Verify at 390px: sidebar does not cover the header. Check the browser console is clean.

Burn list

NEVER is a changelog of last week's mistakes. Keep it specific.

When: The agent keeps repeating a bad habit.
Before

Write clean code.

After

NEVER: new markdown files, extra abstractions for one-time work, rewriting working CSS.

Patterns distilled across 6 products. Use them to write your own roles — don't paste another lab's hidden prompt into a different model.