Write the specs

What are you writing?

One line is enough. Step 1 asks the rest in your AI app, and nothing is done until you approve the brief it writes.

In the case folder's materials/: The notes, documents or past examples it should rest on.

The 5 steps

1. 01

   Gather and frame

   Then you approve, before the next step starts.

2. 02

   Outline

   Then you approve, before the next step starts.

3. 03

   Draft

4. 04

   Independent check

5. 05

   Finish

How it does this well

One writer from the person's material, checked by a fresh reader. Not found and out of scope are valid answers here: it says so instead of guessing.

How the field does this well

• Write the check first: a failing test, a typecheck or a command that comes back pass or fail.
• Work one unit at a time and show the command that ran and its output as evidence.
• Have a reader who did not write the change review it against the written spec, and skip style.
• Keep state in files: a list of checks whose status may only flip to passing, and a progress note.

Ground it first

Ask for, or look in materials/ for: the repository or files to change, how its tests run, and the coding standards. If there is none, say plainly that the result rests on general defaults, and name them.

What stays with the person

Scoping ambiguous work, architecture choices, and the final merge or release. Prepare these; never decide them.

The check that makes it trustworthy

A failing check written first, a passing run shown as output, and an independent review of the change against the spec.

The person's own data

If the person has connected GitHub, Linear, Jira + Confluence, Postgres, read their data through that connector instead of asking for exports, and say which source each number or fact came from. Without one, work from materials/.

Also from finance

• Detect variances, filter by materiality stated up front, explain each with the entries behind it, and assign an owner.
• Tie out totals in code, never by the model.
• Match with deterministic rules first and leave the rest unmatched for a person; never force a match.
• Draft the narrative from plan, actuals and last period's commentary; the person supplies causality.
• What stays with the person here: causal explanation of material variances, booking or approving entries, and sign off.
• The check here: totals tie out in code, every explanation cites the entries that produce it, and unmatched items are listed.

The agents

checker

Checks a Cheatcode draft against its brief and its claims with fresh eyes, pass or fail per criterion.

Where things stay

On your computer

Your files in materials/, every result in outputs/, the work itself. It runs in your AI app, on the plan you already pay for.

With Cheatcode

The steps, the gates and how far each case has come, so a case survives closing the app. See your cases.