Acceptance checks before code
Write down what will prove the outcome while you plan, before any code exists. Each check names the kind of evidence that shows it.
At Blume, we decide how an outcome will be checked before we build it. The checks go in the plan, each a plain statement of something that must be true and the kind of evidence that will show it.
What goes wrong without it
Tests after code
The outcome is “A person refunds an order from its page, fully or in part.” The agent builds it, then writes tests for what it built. All green.
In review you try a partial refund of 60 on an order with 50 left, and it goes through. Nobody had said a refund must not exceed what is left, so the agent did not build that rule or test for it.
Checks first
The checks are written in the plan first:
- A refunded order shows its refund and the amount left. End to end.
- A partial refund never exceeds what is left. Unit test.
- The payment provider records the refund. Integration test.
- The refund email reads well. Checked by hand.
The second check fails on the first run, and the agent adds the limit. At review, each line has its evidence beside it.
What to do
- State the result, not the test. “A partial refund never exceeds what is left”, not “Add tests for refundMath”.
- One thing per check, and name the evidence that fits it.
- Cover the edges you know: empty states, limits, permissions.
- Use the person’s words. An agent that sees a gap asks. It does not invent requirements.
- Keep them few. Four good checks are read. Twenty are skimmed.
In Blume
- Acceptance checks are a section of the plan. A check is met when evidence of its type passed at the current code.
- Agents pass every check before asking for the outcome review.
- With plan review, you approve the checks before the agent builds.