The 5-part prompt that fixes 90% of bad AI output

The problem with "just ask it"

Most people prompt like they're texting a friend: one line, no context, no format, no constraints. Then they blame the model when the output is mid.

After running PromptForge against thousands of real jobs — cold emails, product briefs, legal letters, meeting summaries — we found the same five components appear in every prompt that consistently works. Miss one and quality collapses.

The 5 parts

1. Role — who the model is pretending to be (a specific one, not "expert")
2. Task — the exact verb and outcome, in one line
3. Context — everything the model needs that it can't reasonably infer
4. Format — the shape of the output (list, table, JSON, headings, length)
5. Constraints — what to avoid, what to include, tone, audience, examples

Before vs after

Before (a typical one-liner):

"Write me a cold email to a marketing director about our SEO tool."

You'll get an email that sounds like every other AI email — generic hook, three bland benefits, "let me know if you'd like to chat!" close. Deletable in one second.

After (5-part):

Role: You're a founder writing to a peer, not a salesperson. First-person, direct, no hype.

Task: Write a 90-word cold email to a Head of Marketing at a mid-market B2B SaaS asking if they'd try our SEO tool for a specific outcome.

Context: Our tool finds pages already ranking on page 2 of Google and rewrites them for the top 3. Average lift on tested pages is 42% more organic clicks in 6 weeks. We're free for the first 10 pages.

Format: Subject line (max 6 words, no emojis, no colons). Preview line. Email body. One-line P.S. with a specific ask.

Constraints: No "hope this finds you well", no "quick question", no exclamation marks. One concrete number in the body. Assume they're busy — respect their inbox.

Why this works

Each part removes a class of failure:

• Role kills the "I'm an AI assistant" register
• Task stops the model from padding or hedging
• Context stops the model from inventing facts
• Format eliminates the "should I use bullets or paragraphs?" wobble
• Constraints trains the model on the specific voice you actually want

Steal this template

Role: [Who, in one sentence, being specific]
Task: [Verb + outcome + who it's for]
Context: [Facts, numbers, product truth, links]
Format: [Shape, length, sections, examples]
Constraints: [Avoid list + must-include list + tone]

Drop that into ChatGPT, Claude, Gemini or Copilot and the quality lift is immediate — usually the difference between "unusable" and "I'd actually send this".

---

PromptForge auto-builds all five parts from whatever rough idea you type. Try it free →