Write done as a script

Theory: The four parts of a loop and the goal test build page. "Improve the error handling" cannot terminate a loop, because nothing can ever say it is finished. The goal must be phrased so that a program — not a person, not the model — returns true or false against it.

Workflow

1. Extract the claim. Ask what the user wants true at the end, then restate it as verifiable facts: which command exits 0, which file exists, which string appears, which number crosses which threshold. If a part cannot be checked by a program, say so and negotiate it down to what can.
2. Prefer checks that already exist. The test suite, the linter, the build, the type checker. A goal test that shells out to make test is better than one that reimplements it.
3. Write goal-test.sh (or .py if the checks are easier there) in the project root: runs every check, prints one line per check with pass/fail, exits 0 only when all pass. Keep it under ~40 lines; it must run in seconds and be safe to run repeatedly.
4. Run it now. It should fail before the work is done — a goal test that passes on the current state is testing nothing. Show the failing output.
5. Offer the loop. If the user wants the agent driven until done, wrap it:

#!/usr/bin/env bash
# goal-loop.sh — run a headless agent until the goal test passes or attempts run out.
set -u
MAX_ATTEMPTS=5
for i in $(seq 1 "$MAX_ATTEMPTS"); do
  FEEDBACK=$(./goal-test.sh 2>&1) && { echo "done in $i attempt(s)"; exit 0; }
  claude -p "Goal: <the goal>. The goal test currently fails with:
$FEEDBACK
Fix the code so the goal test passes." \
    --permission-mode acceptEdits --output-format json > ".attempt-$i.json"
done
echo "goal test still failing after $MAX_ATTEMPTS attempts — falling back to a human"
exit 1

Adjust the agent command to whatever CLI the user runs. Keep the three brakes visible and named: the checker outside the model (goal-test.sh), the stop rule (test passes), the budget (MAX_ATTEMPTS, plus a dollar cap read from the JSON output if they want one).

Rules

• The checker never asks the model's opinion; self-review is not a checker.
• Feed the checker's output back verbatim — the failure text is the prompt.
• If the user cannot state done as a check, they do not have a loopable task yet; tell them that plainly and suggest an interactive session instead.