← AI Library
AI Systems · WorkflowClaude

The Double-Decker AI Sandwich

A simple way to place human judgment between AI passes, then decide which checks can later become rules, examples, skills, or evals.

Todd Kelley · Rewritten field guide from a public Cindy Zhu source

WorkflowClaudeApprovalEvals
Approval Loop Prompt 
Turn the workflow in this conversation into an AI approval loop.

Put human judgment after any step where truth, safety, taste, intent, or an irreversible action can drift.

Use objective checks for facts, files, calculations, links, and system state.
Use human approval for taste, tradeoffs, judgment, stakes, or anything consequential.

Return the workflow as:
human sets outcome and standard -> AI pass -> check or approval -> AI revision -> final verification.

For every checkpoint, state the evidence required and what happens if the check fails. Then propose which repeated corrections should become rules, examples, skills, or evals.
Sorting Rule

Objective Check or Human Call?

Objective Check

There is a right answer someone or something else can verify: a link resolves, a number matches the source, a file exists, a test passes.

Human Call

There is no single correct answer: tone, taste, the right lead, whether the work sounds like you, or whether a tradeoff is acceptable.

Graduation Ladder

How a Checkpoint Gets Automated

  1. Start with a rule. Add the correction to the prompt or project instructions.
  2. Add examples. If the rule keeps being misunderstood, show a good and bad version.
  3. Turn it into a skill. Save the whole repeated process once it becomes multi-step.
  4. Build an eval. Use tests only for checks with evidence and a clear pass/fail condition.
  5. Remove one gate at a time. Only retire checks after repeated real-world passes.

Never remove every human gate. Health, legal, finance, security, privacy, public publishing, and irreversible actions keep a person in the loop.

Original Source

Adapted from Cindy Zhu's guide: Prompt to add human approval checkpoints.