1. Name the check, and make the session run it
The test command belongs in AGENTS.md, so you are not inventing it per task. Then the brief says to run that command. “Done means the handbook’s test command passes” is only a finish line if that command actually runs.
Grok can run that command. The scrollback shows the command and the output. Read the output. A paragraph that says “all tests passed” is the model’s summary of the output. The output is the evidence. If the command was not the one in the handbook, the green line is about a different check.
2. You read the diff
Edits show up as diffs in the session. Open them. Generation is fast, and a wrong line is quiet. Look for four things:
- Files you did not expect. A drive-by rewrite of a neighbor is a bug even when the tests pass.
- Behavior you did not ask for. Extra options, extra dependencies, a new framework.
- A deleted check. A test removed to make the suite green is not a fix.
- A secret, or a real customer value, that landed in a file.
If you cannot explain a hunk, ask in the same session: “Why did invoice.py change? Quote the requirement.” Then decide. Asking is not a delay. Keeping an unexplained hunk is how a codebase rots.
A diff that deletes the old invoice test and adds a smaller one is not a pass, even when the new command exits without an error. The deleted test was checking something. Ask why it left. If the reason is “it was failing,” the failure is the work. Deleting the check is not the fix.
3. A fresh session is a colder reviewer
The session that wrote the code is a poor judge of it. It remembers the intent, and it will defend the intent. Start a new session, point it at the diff, and ask for a ranked list of problems with the evidence for each one. Tell it not to edit.
Review the uncommitted diff.
Rank real problems first.
For each one, name the file and why it is wrong.
Do not edit anything.
You still decide. A review session can be wrong too. It is useful because it does not share the writer’s momentum. Do not merge, or ship, a change you could not explain to a teammate after that review.
4. When you change a skill or a standing instruction, retest old tasks
A code change has a test suite. A change to a skill, a project rule, or a prompt you reuse does not. Keep a short list of real tasks you already know the answer to. Ten is plenty. After you edit the skill, run a few of them and compare the result to the last time.
That list is an evaluation. It is yours. Grok Build does not ship it for you. “It feels better” is not a result. One task that used to work and now wanders is a result. Add the failure to the list so the next edit has to face it.
5. A failed run still has a first broken step
If the tests fail, go back to the trace before you rewrite the brief. Which command failed. Which assertion. Which file the edit actually touched. Point Grok at that step. A new adjective on the original sentence throws away the evidence you just paid for.
The last page puts these habits in order, as one workflow for the next real task.