GUIDE

Make your coding agent prove it's done

"Done!" from an agent means nothing until you've seen it run. Here's how the Claude Code team handles that, cut down to a checklist you can use alone. Sourced Oct 20, 2026.

What the Claude Code team reported

In an Oct 15, 2026 post, Anthropic said Claude writes about 90% of Claude Code's own code. From June to September, merged pull requests more than doubled while the share that had to be reverted fell by more than 40% (source). They didn't stop reviewing. They moved most of the checking to automated layers and saved human attention for risky changes.

The first layer is the one you can copy today: give the agent something it can run, and make it show the output instead of saying it's done. Boris Cherny, who created Claude Code, estimates that giving Claude a way to verify its work makes the final result two to three times better (quoted in the same post).

The solo checklist

  1. One check command. Put a single command in your rules file that runs lint, tests and a build. The agent runs it after every step.
  2. Show, don't claim. Tell the agent its work isn't finished until it pastes the check output, a screenshot, or a test result. "It should work" doesn't count.
  3. Fresh-eyes review. Before you merge, have a second agent session (or a subagent) review the diff with a clean context. It catches what the author session talked itself into.
  4. Fix it at the source. If review flags the same mistake twice, add a line to your rules file or a lint rule so it can't come back.
  5. Scale by risk. Docs and copy changes can ship on green checks. Anything touching auth, payments, saves or deploy scripts gets your own eyes, every time.
  6. Make undo cheap. Small commits, a branch per task, and feature flags for anything big. A fast rollback is cheaper than a perfect review.

Copy-paste rule for your agent

Before you say a task is done:
1. Run ./scripts/agent-check.sh and paste the last 20 lines of output.
2. If it's red, fix it. After 3 failed fixes, stop and tell me what you tried.
3. Commit only on green, with a message that says what you verified.

Making a game?

Games add a check that tests can't run: is it still fun? Pack #001 ($19) adds a playtest agent with tiered PASS/FAIL checks and an "Is It Fun?" review on top of the same loop.

Get the loop for free

Our free Checkpoint Loop is a rules file plus agent-check.sh, one command that auto-detects npm, pytest, Cargo and Go and blocks obvious secrets in your diff. It does steps 1 to 3 out of the box. See also: one rules file for Claude Code, Codex and Cursor.

Get the free Checkpoint Loop

Get the patch notes free →

Free, weekly, one-click unsubscribe. We note changes like this one every week.