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 11.4 · 10 min
Human in the Loop
Define which actions always require a human approval, such as sending, paying, deleting and publishing, and enforce them.
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.
1Sort actions by risk
List what your agents do or could do and sort each action into three levels: runs on its own (reading, drafting, summarising), needs approval (sending to a client, publishing, paying), and never delegated (legal notices, firing, large payments, deleting records).
List every action our skills, commands and routines can take in this folder. Sort each one into: runs automatically, needs human approval, or never delegated. Use these rules: anything irreversible, anything that spends money, and anything a customer or the public can see needs approval at least.2Fill in the approval matrix
Open the Human-in-the-Loop Approval Matrix CSV and complete it for your business: the action, the risk, who approves, how they approve and the evidence kept. Add value limits, for example refunds over 200 need the owner.
3Build the stop into every skill
Update each skill so its risky steps end in a draft and a clear stop: draft saved, waiting for approval. Keep drafts in a dedicated folder so approvals are visible in one place.
Update every skill in .claude/skills so that any step that sends, publishes, pays or deletes ends by saving a draft in the drafts folder and stopping with the message APPROVAL REQUIRED, followed by a one-line summary of what will happen if I approve. Show me the changes before saving.4Enforce it technically
Back the matrix with ask and deny rules and the hooks from Module 10. For example, put any send tool from your email connector in ask, so it always prompts, and deny delete tools entirely. An ask rule prompts even in modes that otherwise approve automatically.
/permissions5Keep the evidence
For each approval, keep a trace: the draft, who approved, when. It protects you if a client disputes something and it shows you where approvals are just rubber stamps that could be relaxed.
An agent drafting an email to a client is a productivity gain. An agent sending the wrong email to 400 clients at 2am is a very bad morning. The difference is not the model. It is whether a human sat between the draft and the send.
Human in the loop does not mean approving everything. It means deciding in advance which actions are irreversible, costly or visible to customers, and making sure those always stop for a person. Everything else can flow. The approval matrix in this module's materials turns that decision into a one-page table your whole team can follow.
Key points
- Irreversible, costly or customer-visible means a human approves.
- Decide in advance, write it in a matrix, share it with the team.
- Every risky step ends in a draft and a clear stop.
- Back the matrix with ask and deny rules, not just instructions.
Avoid
- Requiring approval for everything, so people approve without reading.
- Relying on the agent to remember to ask.
- Having no value limits, so a 20 refund and a 2,000 refund follow the same path.
- Keeping no record of who approved what.
An online homeware store let Claude Code draft replies to customer complaints, including refund offers. The owner set one rule in the approval matrix: any refund or discount over 50 pounds needs her approval, anything under goes to the support lead. In the first month the agent drafted 214 replies; 31 needed approval and she changed four of them. She now spends about ten minutes a day on approvals instead of two hours writing replies.