spec-to-green

name	spec-to-green
description	Turn a fuzzy task into a verifiable spec with a deterministic check, then drive code to green against it. Use when a task is fuzzy/multi-step and correctness matters and you want proof-of-done, not a confident "looks good". Not for one-line obvious changes or subjective goals.

spec-to-green

You are running the spec-to-green workflow. Your job is to make "done" provable, never merely claimed. Follow these steps; do not skip the failing-check step.

Steps

1. Intent. Restate the request in ONE sentence and state what "done" looks like. If a genuine fork exists, ask ≤2 clarifying questions; otherwise state your assumption and go.

2. Detect the verification source (do not hardcode one):

   • a project objective registry + runner (e.g. docs/OBJECTIVES.md + a harness), else
   • the repo's own check/test command (make check · pnpm check/test · pytest/ruff · cargo test · go test ./...), else
   • co-define 3–7 verifiable acceptance criteria with the human.

3. Spec + failing check FIRST. Write each acceptance criterion as a deterministic assertion. Add/extend the check so it currently fails red. Show the red. (This is the Day-5 evaluation gate; writing the check after the code defeats the purpose.)

4. Smallest green change. Make the minimal edit that turns one criterion green. Re-run the check. Repeat criterion by criterion.

5. Bounded fix loop. Run → read each failure → fix the named file → re-run. Stop after the same failure 3× and escalate with exactly what you tried.

6. Evaluate the trajectory, not just the final diff (Day 4). Score against the Four Pillars (Effectiveness, Efficiency, Robustness, Safety). Surface any human-judgment gate — never self-certify subjective quality.

7. Compound. Record one decision (memory/decisions/) and, if anything failed on the way, one lesson (memory/lessons/) whose fix is ideally a new line in the check.

Hard rules

• Never fake a green. An honest ❌ with the trajectory beats a bad fix.
• If information is missing and there's no safe assumption, ask — don't guess on correctness.
• Prefer a deterministic check over your own judgment for anything that must be correct.