Write a serious prompt. Check what comes back.

Everything you need to write a prompt that gets real output instead of slop — and the safeguards to catch AI when it lies, flatters, or skews. Keep this open every time the work actually matters.

Part 1 · The Success Brief — 5 fields, every serious prompt

Field	What goes here	Example
Output	What you want delivered. Format, length, type.	"150-word IG caption + 3 hook variations"
Purpose	Why it exists. Who it's for. What it must accomplish.	"Get female founders to DM me about the Bootcamp"
Constraints	The rules. What MUST be true.	"Under 150 words. No em-dashes. No 'unlock.'"
Goal	What success looks like — concretely.	"5 DMs in 24 hours saying 'tell me more.'"
Anti-Goal	What you do NOT want. The taste.	"Don't sound like ChatGPT. Not motivational."

Anti-goals are the taste. The 80% that says what NOT to do is what makes the output sound like you.

Part 2 · The fill-in scaffold — copy, complete, paste

OUTPUT: [format, length, type] PURPOSE: [who it's for + what should happen when they see it] CONSTRAINTS: [rules, banned words, must-haves] GOAL: [what success looks like, concretely] ANTI-GOAL: [what you do NOT want it to sound like] EXAMPLE: [paste one you love] — match this energy, different topic.

The show-don't-tell ruleA good prompt is 100–300 characters WITH an example. Don't make Claude guess your taste — show it. No example handy? Ask Claude to build you an empty template, or to reverse-engineer a piece you love into a reusable style profile.

Part 3 · The Discernment Combo — paste at the bottom of any serious prompt

This one block defeats all three failure modes at once — hallucinations, sycophancy, and bias.

A few rules I need you to follow: 1. If you're not 100% sure something is factually true, say "I don't know" and flag it with [VERIFY]. Do not invent stats, citations, or quotes. 2. Don't tell me what I want to hear. Tell me what a smart skeptic would say. What are the 3 weakest parts of this? 3. After your first answer, argue against yourself. Make the strongest case for the opposite position. Treat me like a CEO who needs the truth, not a client who needs flattery.

Part 4 · The 5-Question Pressure Test — before you ship anything

Is anything here a fact I should verify?

Catches hallucinations.

Is this in MY voice or a generic AI voice?

Catches sycophancy and sameness.

What's the failure mode I can't see?

Catches blind spots.

Did it actually answer what I asked?

Catches AI laziness and pattern-matching.

Would I be embarrassed if this shipped as-is?

The final gut check.

RememberWhatever AI produces under your name is YOUR responsibility. The model doesn't get sued or embarrassed in front of your client. You do. You are the QA.