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 11 · Lesson 11.5 · 10 min
Verifying Agent Work
Create quality checks and a simple sampling routine so you catch errors before customers or partners do.
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.
1Make the agent show its sources
Require every number and every factual claim to cite the file, row or page it came from. Unsourced output is a warning sign on its own.
2Ask for a self-check against a checklist
Give each important skill a short checklist and ask Claude to check its own output before handing it over. It will not catch everything, but it catches a surprising amount.
Before you give me the report, check it against this list and show me the result: every number cites a source file; totals add up; dates are in the right month; no client names from other folders; nothing marked as fact is actually a guess. Fix anything that fails and tell me what you fixed.3Use a separate reviewer
For high-stakes work, have a fresh session or a reviewer subagent (Module 7) check the output with no knowledge of how it was produced. A reviewer that did not write the work is much better at spotting its errors.
You are a sceptical reviewer who did not write this report. Compare reports/2026-09/monthly-report.md with the source files in reports/2026-09/data. List every number that does not match its source, every claim without evidence, and anything that would embarrass us if a partner read it.4Sample and log errors
Decide a human sampling rate per task: 100 percent for the first two weeks of any new skill, then for example one in five for routine drafts and every headline number in reports. Log each error you find in a simple list with the date, the skill and the fix.
5Turn errors into rules
Every error you find should change something: a line in the skill, a rule in CLAUDE.md, a new check in the checklist or a hook. The error rate should go down month after month. If it does not, the skill is not ready for more autonomy.
Here is the error log for the last month. Group the errors by cause and propose the smallest change to our skills, CLAUDE.md or hooks that would prevent each group from happening again.Agents are confident. That is useful and dangerous at the same time. A report with one wrong number looks exactly as polished as a report with no wrong numbers, and a polished error travels further than a messy one.
Verification is not about distrusting the agent. It is the same thing you do with a new employee: check everything at first, then check a sample, and increase the sample the moment something slips. You will set up three layers: the agent checks itself, a second agent checks the first, and a person checks a sample.
Key points
- A polished error travels further than a messy one.
- Three layers: self-check, separate reviewer, human sample.
- Check everything at first, then sample, then adjust.
- Every error found should become a rule.
Avoid
- Trusting output because it looks professional.
- Letting the same session that wrote the work also review it.
- Reducing checks after one good week.
- Fixing errors in the output instead of fixing the skill.
A US employment law practice used Claude Code to draft first versions of client update letters from case notes. In the first two weeks, a paralegal reviewed every letter and logged nine errors, mostly dates copied from the wrong matter. They added a self-check comparing every date against the case file and a rule to work in one matter folder at a time. Errors in the next month dropped to one, and review moved to a one-in-four sample plus every letter mentioning a deadline.