Free tool · The Done Check

Agent work needs a closeout check.

Your agent says it’s done. Make it prove it.

Your coding agent just reported the task finished — and you’ve been burned before: last time “done” meant a test that never ran, the wrong file edited, or a passing result it made up. Paste the prompt below the moment it claims “done.” It makes the agent show its own proof instead of a confident summary.

Free, no email required · Works in Claude Code, Cursor, Copilot, or any AI coding agent

Paste this to your agent

The Done Check

the-done-check
Before you tell me this is done, run this check on your own work. Answer every line. If you can't answer one, the verdict is NOT DONE — say so instead of guessing.

1. ARTIFACT — Point to the exact thing that changed: the file path, route, command, or record. The location I can open, not a description of it.

2. PROOF — Run the smallest thing that exercises that artifact and paste the REAL output: the test and its result, the command and what it printed, the page you loaded, the query and its row count. A summary of the output is not the output.

3. MATCH — Restate what I asked for, then what you produced, side by side. If they differ at all, name the difference.

4. BOUNDARY — List anything risky you touched: files deleted or overwritten, credentials or config changed, anything sent to a live system. If none, write "none."

5. VERDICT — Pick one, nothing softer:
   - DONE — artifact shown, proof pasted, matches the ask, no unreviewed risk.
   - DONE, WITH RISK — it works, but name the risk and who should check it.
   - NOT DONE — what's missing, and the single next step to close it.

Rules: no "should work," "looks good," or "I think." Every claim needs an artifact or a real command output beside it. If showing the proof would take longer than the task did, show it anyway.

Keep it in a snippet or scratch file. Fire it whenever an agent says a task is finished, then don’t merge, publish, or send until the verdict comes back DONE.

Proof before download

Five answers, each with something you can open.

The gate does the distrusting so you don’t have to re-check the whole task by hand. An agent that has to paste its own proof can’t hide a half-finished job behind a summary.

  1. ArtifactThe exact file, route, command, or record that changed — a location you can open.
  2. ProofThe real output of the smallest thing that exercises it, pasted, not summarized.
  3. MatchWhat you asked for next to what it built, with any difference named.
  4. BoundaryAnything risky it touched — deletes, credentials, live sends — or an explicit “none.”
  5. VerdictDONE, DONE WITH RISK, or NOT DONE — no “should work” allowed.
When to fire it

Best right before something ships.

Use when the agent claims a page changed, a workflow shipped, code is ready to merge, data moved, or a message is ready to send — anything you’d have to trust without re-checking.

Skip when it’s loose exploration, throwaway scratch work, or research you expect to stay unresolved — there’s no artifact to prove yet.

Weak close

“Done. Build passed. I think the page is fine.”

Strong close

“Changed /writing/x/; build and screenshot shown; no secrets or sends; you own the launch.”

Prefer it on paper

Grab the one-page card.

Same gate, laid out as a single branded page you can pin next to your desk or drop in a team wiki. The prompt is printed in full, ready to retype or reference.

Download the PDF card