Skip to main content
When a Claude agent finishes, the task moves to Testing. The review system lets you evaluate the work and either approve it or send it back for revisions.

​
Review Actions

Click Approve to mark the task as done. The task moves to the Done column and no further agent work happens.

​
Feedback Loop

The revision cycle can repeat as many times as needed:
1

Claude finishes → Testing

Review the changes in the terminal log or your IDE.
2

Request changes with feedback

Describe what’s missing or wrong. Be specific — Claude performs better with clear instructions.
3

Claude iterates → Testing

The agent picks up your feedback, makes adjustments, and returns to Testing.
4

Approve → Done

When satisfied, approve the task.

​
Revision History

Every revision is recorded. The task detail view shows:
  • Each round of feedback you provided
  • The agent session logs for each iteration
  • Token usage per revision
  • Total cumulative cost
Write feedback as if you’re doing a code review. Bullet points with specific file or function references help Claude make targeted fixes.

​
Best Practices

Instead of “this doesn’t work,” say “the /api/users endpoint returns 500 when the email field is missing.”
Check the git diff or terminal logs. Claude may have made changes beyond what you expected.
Large tasks with vague descriptions lead to more revision cycles. Break work into smaller pieces.