Split an overloaded prompt into a checked three-step chain

One prompt is asked to plan, draft, and check itself, and it quietly skips one of those jobs.

Beginner to intermediateRead 6 minDoing: an hour with your API key and an editor open7 steps

▶ Start

What you'll have at the end

A three-step pipeline: draft, gate check, and one automatic rewrite when the check fails

Before you start

You need

One live prompt that already asks a model to decide an action, write a customer-facing reply, and confirm the reply follows a policy rule, all inside a single call, plus a working setup for sending a request to your LLM API and getting a structured JSON reply back.

Not covered

This does not cover a policy rule that can't be reduced to comparing two fields, or a gate that needs to branch into several different recovery paths: those need a second LLM call judging prose, or a routing pattern, neither of which this fixed three-call chain does.

Leans on

Find the seam, then split the draft off

Name the clause your prompt skips
Give the draft call one job
Run the draft call and keep its fields

Build the gate and catch the real failure

Write a gate that checks the number
Run the gate and read the fail

Fix it once, and stop

Feed the exact reason to a rewrite
Re-run the gate and cap the retry

Where to go next

Your agent never checks its own answer before handing it back

for a checker that reads prose against real tool data with a second LLM call, when the policy rule can't be reduced to a single field comparison the way this recipe's gate is.

Run three independent LLM checks in parallel instead of one after another

when the checks you need don't actually depend on each other in order, since running them at once beats paying for three sequential calls only one of which usually needs to run.

Checked 18 Aug 2026

Part of the AI Agents cookbook