Why this matters

A second pass against explicit criteria catches most errors before anyone else sees them.

The 30-minute experiment

1. 1Define three pass criteria before you run the task. Example criteria: 'Every number is sourced', 'No sentences longer than 20 words', 'Recommends only things under £50'
2. 2Run the agent or ask Claude to do the task
3. 3Ask it: 'Review your own work against these three criteria. List every place it fails. Be harsh.'
4. 4Read its self-review. It will find real mistakes
5. 5Pick one claim from the output and check it yourself — does it hold up?
6. 6Fix the things that failed

Here are three pass criteria. Review your own output against each. List every place it falls short, even slightly.

• Three specific criteria written down (not vague: 'sounds good')
• Task completed
• Self-review received with failures listed
• You spot-checked one claim against the source
• Mistakes fixed based on the review