Agent work is reviewed before it reaches Simon

Context

On 12 November, my agents built about fifteen pieces of work. I checked each one before passing it to Simon, and the checks found something most times:

• an audit that would have published a map of security weak spots
• a brief that was wrong about the code it described
• credits to other people’s work that the agent had not checked
• a test that proved nothing

None of those checks was written down anywhere. They lived in the habits of one Claude session. A new session, a different AI model, or an agent that Simon ran directly would have skipped them. It is the same problem as a small business whose process lives in one employee’s head.

Options

1. Trust the reports. Fast, and wrong often enough to matter.
2. Automated checks only. The automated checks (CI) already catch mechanical problems, such as em-dashes, badly formed journal entries and broken links. They cannot tell whether a change did what was asked, looks right on a phone, or credits the right author.
3. Simon reviews everything in detail. Simon is the final approver, but reading every line of every change would turn Simon into the bottleneck and defeat the point of agents.
4. A written review step, followed before anything reaches Simon. A skill sets out what to check for each kind of work, and CLAUDE.md makes it a rule.

Decision

Option 4. A skill, review-agent-work, holds the judgement part: treat a report as a claim, check it against reality, and use a checklist for code, website changes, copy, research and the journal. It also covers merges and what follows them, instructions found inside the work, and how to report back to Simon, including the reviewer’s own mistakes. CLAUDE.md makes the step a rule. CI keeps the mechanical checks, and the review makes sure CI actually ran.

Review is a separate step from building. The agent that did the work is never the only one who checks it.

Consequences

• Work takes a little longer to reach Simon, and arrives with a short account of what was checked and what was not.
• The review step now works the same in any session, and can be improved in one place.
• It is different from the agent safety rules (issue #13). Those decide what agents are allowed to do, and stop harm before it happens. Review catches what got through.
• The skill lives in the Roofus repository while I refine it. Once it has settled, it moves to method7/house-rules, so every Method7 project uses it.
• Simon agreed the approach on 12 November 2026.