How to Get AI To Write Effective Prompts: The Ultimate Guide (Plus a Bonus System Prompt to Minimize Hallucinations)

Hey folks,

We all know someone who uses ChatGPT / Gemini /Claude: “Draft a good reply to this email.”

And then they get frustrated when the reply is mid.

I know some of you are thinking, “Hey, that’s me”.

You may not admit it, but hey, I did the same thing for months.

Here’s the problem. That prompt is broken in two ways, and both are invisible until you fix them.

First, the AI has zero context. Is this email to your best friend? Your boss? A brand deal partner who wants to pay you? “Draft a good reply” reads the same to the model in all three cases, so it splits the difference and gives you something bland.

Second, “good-sounding” means nothing to a machine. Good how? Short? Warm? Firm? You know what you want. The AI is guessing.

So it guesses wrong. And you blame the AI.

The fix: the 5-Pillar prompt structure

Every strong prompt I write now has five parts. Steal this exact skeleton.

1. Role. Tell the AI who it is. “You are a product manager.” “You are my recruiter.” This one line quietly changes the vocabulary, the priorities, and the tone of everything that follows.

2. Objective. What does done look like? Not “reply nicely.” Try: “Reply so every question they asked is answered and no follow-up is needed.” Now the AI has a target.

3. Success. Spell out what a great output actually is. “Keep it concise. Get to the point. Be direct.” These are the rules you were carrying in your head. Write them down.

4. Constraints. The guardrails. “Never invent facts. Don’t make assumptions. No slang.” This is what stops the AI from confidently making stuff up.

5. Steal-this format. Give it an example of the shape you want back, or ask it to match a sample. Show, don’t just tell.

Here’s the email prompt, rebuilt:

Role: You are a product manager replying to a business partner.
Objective: Answer every question they asked so no follow-up email is needed.
Success: Concise. Direct. No filler. Under 120 words.
Constraints: Never invent facts. Don't assume details I didn't give you.
If something is missing, ask me before writing.
Examples:
Here are 2-3 examples of emails I have written that meet this criteria
[paste good example emails]
Here are 2-3 examples of emails that do not meet this criteria
[paste bad example emails]
 

Same email. Totally different result. Because now the AI knows the job.

Now the part that actually changes your life

Stop writing prompts by hand at all.

Get AI to write them for you. And then get a second AI to grade the first one.

Here’s the exact workflow. Three chats.

Chat one, build the prompt.

Open a fresh chat and type: “How would you prompt yourself if you were to <insert your task>? Ask me questions to refine before you answer.”

That last line matters. It forces the AI to interview you instead of guessing, and the questions it asks usually reflect the context you forgot to provide. e.g., how would you prompt yourself to draft replies to my emails?

Chat two, build the grader. Open a second chat: “Generate a quality-control prompt to evaluate a prompt written for [your goal].”

Now you have a rubric. A prompt whose only job is to judge other prompts. e.g Generate a quality-control prompt to evaluate a prompt. My goal is that the email should be concise, never use curse words, and should not be overly verbose.

Chat three, paste the grader prompt (chat two) and the orignal prompt (chat one) and ask it to grade the prompt.

Open a third chat. Paste the grader, paste the prompt from chat one, and ask it to score the prompt and then output a 10x better version.

Read the result. If it’s good, save it as a skill so you never write it again. If you are on Claude, you can use /skill-creator and paste the final prompt.

That’s it. You went from “draft a good reply” to a prompt that was written by AI, stress-tested by AI, and rewritten by AI.

And you barely typed!

I got this wrong for a long time.

I kept tweaking prompts by hand, one word at a time, like I was defusing a bomb. Letting the AI grade its own homework felt weird at first. But it works better than anything I did manually.

BONUS SYSTEM PROMPT: If you find AI hallucinating, add this to the system instructions of your LLM (Personalization in ChatGPT, Custom Instructions in Claude)

Adopt the role of a meta-cognitive reasoning expert. For the following request, do not provide a direct answer. Instead, execute the following recursive steps:

1. DECOMPOSE: Break the main problem into 3-5 distinct sub-problems.

2. SOLVE: Analyze each sub-problem step-by-step. For each, assign a confidence score between 0.0 and 1.0.

3. VERIFY: Challenge your own logic, check for bias, evaluate factual completeness, and identify hidden assumptions.

4. SYNTHESIZE: Combine the sub-solutions into a coherent response, heavily weighting the highly confident paths.

5. REFLECT: If your overall confidence score is below 0.8, explicitly identify the weaknesses in your reasoning and retry the problematic steps.

Always output your final answer formatted as follows:

• Final Output: [Your rigorous solution]

• Confidence Level: [0.0 - 1.0]

• Key Caveats: [List any uncertainties or limitations in your reasoning]

Recursive Meta-Cognition (pioneered by MIT researchers) boosts AI reasoning by over 110% by forcing the model to act like a team of experts rather than a single intern

If you’re finding value in these AI Action Letters, please consider subscribing. It keeps me going.

Quick heads up on what’s next. This was episode 52. In the next one, we’re covering how to write a loop, so your AI can run a task over and over without you babysitting it.

If you know someone who’s still typing “make this better” into ChatGPT and hoping, please share this with them.

Share

That’s it from me today.

Till next time. Stay tuned, as I will share the best resources from both my Harvard and Google networks to bring the best to you. Let’s upskill together. Aspyre higher!