# Build a human approval queue for AI-assisted work

Original 5HO practice material. Fictional example. Reviewed 5 October 2026.

## Useful result

You'll finish with a visible queue that keeps a human accountable for execution.

Drafting and approving are different jobs. Put a visible gap between them, so a helpful assistant cannot turn “looks promising” into a sent message. An approval queue can start as a plain table.

## Set up

Allow 25 minutes. Use a spreadsheet or document and a chat tool. No live sending integration is needed for the first version.

## Practice input

Queue items: draft venue query, owner Jo; proposed AUD 50 credit, approver not assigned; draft training reminder, owner Lee. Status choices: Draft, Needs review, Approved, Sent.

## Do the task

1. Create an item ID, draft text, reviewer, status and decision note for each item.
2. Ask AI to prepare drafts and put all three into Needs review, without executing them.
3. Check that only a real reviewer can record approval and that approved content is not silently changed afterwards.

## Prompt

Build an approval queue from these items. Leave the credit approver Unknown. Draft messages only. Include the exact version reviewed and a field for the human decision; do not mark anything Approved.

## Refine

Add a rule that changing approved wording returns the item to Needs review. Separate approval from the later send action.

## Expected checks

No row should be Sent. The unassigned credit cannot move forward until the proper authority is identified. A status word in a spreadsheet does not itself enforce an API permission.

## Save

Save approval-queue.csv and test one approve/revise cycle manually before connecting a sender.

## Sources

- Tool permission controls in Claude Code: https://code.claude.com/docs/en/permissions