MODULE 14 / 18 · PART 5 · LEAD THE AGENT & THE INTERFACE
AI-Agent Meta-Prompts
Improve every other prompt by changing how you direct the agent.
THE SHORT ANSWER
What this module changes.
These prompts improve every other prompt because they change how you direct the agent. Make it plan, state assumptions, cite files, stop before guessing, and leave a handoff.
These prompts are about driving the agent itself. They are the highest-leverage prompts in the book because they improve every other prompt you write.
Long tasks, expensive tokens, dropped connections and shared laptops make agent discipline practical, not theoretical. A five-line handoff at the end of a session is often more valuable than another hour of context.
THE MODULE PROMPT LIBRARY
20 prompts for agent meta-prompts.
Copy any prompt, replace the [bracketed] slots with your own project context, and attach the files the agent needs. Use “Copy all prompts” when you want the full module in one place.
20 copy-ready prompts. Bracket text is a slot for your own context — replace it before you send the prompt.
Before you write anything, ask me the questions you need answered to do this well. Do not start until I answer.
Why it worksPrevents the agent from filling gaps with plausible inventions.
Restate the task in your own words and list your assumptions. Wait for my confirmation.
Why it worksA cheap checkpoint that catches misunderstanding before any code exists.
Work in steps. After each step, stop and tell me what you did and what you plan next.
Why it worksKeeps a long task reviewable and interruptible.
Tell me what you are uncertain about in this task, and what you would need to check to be sure.
Why it worksSurfaces the agent's own confidence gaps, which is where errors hide.
Do not write code yet. Give me the plan, the files you will touch, and the order.
Why it worksA plan is reviewable in a minute; a wrong implementation costs an hour.
If you are about to make an assumption about the codebase, stop and ask instead.
Why it worksAssumptions about unseen code are the top source of broken output.
Show me the diff and nothing else. I will ask for the explanation if I want it.
Why it worksRespects the reviewer's time and keeps the signal high.
What is the smallest change that accomplishes this? Propose that first, before any larger refactor.
Why it worksCounters the agent's bias toward rewriting.
Cite the specific files and lines you are relying on for each claim.
Why it worksMakes verification possible and exposes invented context.
If this task is ambiguous, tell me the two most likely interpretations and ask which one I mean.
Why it worksTurns ambiguity into a question instead of a coin flip.
Review your own output as a hostile reviewer would. What is wrong with it?
Why it worksSelf-critique catches a surprising share of errors before you do.
Do not summarize what you just did. Tell me what is left to do and what could still go wrong.
Why it worksRedirects the agent's verbosity toward useful information.
Keep a running list of the decisions you have made and the alternatives you rejected.
Why it worksProduces the rationale you need for the PR description.
When you are unsure between two approaches, present both with the trade-off instead of choosing silently.
Why it worksSilent choices are the ones you discover too late.
Stay within the scope I gave you. If you see something else worth fixing, list it separately.
Why it worksPrevents scope creep in the diff.
Tell me what you would need to see to be confident this works, and whether you have seen it.
Why it worksDistinguishes verified from plausible.
Explain your reasoning before the conclusion, not after.
Why it worksReasoning-first output is easier to audit for errors.
If you cannot do this reliably, say so and tell me what would make it possible.
Why it worksAn honest refusal is worth more than confident nonsense.
Summarize the state of this task in five lines so I can hand it to another engineer.
Why it worksProduces a handoff artifact at any point in the work.
What is the one thing I should check before I trust this output?
Why it worksDirects your limited review attention to the highest-risk part.
PRACTICE · 20–30 MINUTES
Try it on a project you actually own.
Start your next task with prompts 1, 2 and 5. Do not allow code until the plan is confirmed, then record the decisions the agent made.
- Pick one prompt from this module that matches a real task.
- Replace every bracketed slot with context from the feature you are working on now.
- Attach only the files the agent needs. Keep secrets, credentials and personal data out of the chat.
- Ask for the artifact — a spec, diff, test, runbook or decision note — and the evidence that it works.
- Record what changed and what remains unverified in the repository.
Want the same prompts across every module? Search the full prompt library or return to the course overview.