This lesson is being recorded
Every video of this course is being uploaded to my YouTube channel. Subscribe so you do not miss it.
Bruno Ashford @brunoashfordSubscribeModule 7 · Lesson 7.3 · 11 min
The Reviewer Agent
Put a reviewer agent between your AI's work and your clients, so every output is checked by a fresh pair of eyes.
Use ← → to move. The marker highlights text and the pencil draws on the slide; navigation locks while you annotate. Undo removes the last mark, Esc exits.
1Keep the reviewer read-only
Give the reviewer read and search tools only. If it can edit, it will start rewriting instead of reviewing, and you lose the separation of duties. Its only output is a report: pass, or a numbered list of issues with the exact location of each.
2Give it a real checklist
The reviewer is only as good as its checklist. Include facts versus sources, client name and details, numbers and totals, promises you cannot legally or commercially make, tone versus your brand voice, and anything your CLAUDE.md marks as forbidden.
3Build the write, review, fix loop
Ask for the draft, then the review, then the fix, then a second review. Two rounds catch almost everything. Do not skip the second review, because fixes introduce new mistakes.
Write the proposal for the client in notes/call-acme.md using my proposal skill. Then use the reviewer subagent to check it against the proposal checklist. Fix every issue the reviewer finds, then run the reviewer again. Show me the final version and both review reports.4Grade issues by severity
Ask the reviewer to label issues as blocker (wrong client, wrong price, false claim), important (tone, missing section) or minor (wording). Anything with a blocker never leaves the building, no matter how tight the deadline.
5Encourage automatic use
Put the words 'use proactively' in the reviewer's description, for example 'Use proactively before any document is sent to a client.' That encourages Claude to call the reviewer on its own at the right moments. Still, for anything client-facing, ask for it explicitly until you trust the habit.
The most expensive AI mistake I have seen in a small business was not a hallucinated fact. It was a proposal sent with the wrong client's name in paragraph three. The writer was fine. Nobody checked.
A reviewer subagent is the cheapest insurance you can buy. It did not write the draft, so it has no attachment to it. It reads with a checklist, flags problems, and never rewrites the work itself. Writing and checking are different jobs, and they should be done by different agents.
Key points
- Writing and checking are different jobs for different agents.
- The reviewer is read-only and returns a report, never a rewrite.
- Two review rounds: fixes create new mistakes.
- Blockers never leave the building.
Avoid
- Giving the reviewer edit tools so it silently rewrites the work.
- Using a generic 'check this' instruction with no checklist.
- Skipping the second review after fixes.
A small employment law firm uses Claude to draft first versions of client letters. Their reviewer subagent checks four things: the client name and matter number match the file, every legal claim points to a source in the matter folder, no outcome is promised, and the tone matches the firm's letter style. In the first month it flagged 11 blockers across 140 letters, mostly wrong dates carried over from templates. Each of those would have reached a client.