How to Write an AI Prompt That Actually Works
A practical framework for turning vague requests into precise, testable prompts for ChatGPT, Claude, Gemini, and other AI tools.
The real problem is missing decisions
Weak prompts are rarely weak because they are short. They fail because they leave important decisions to the model: who the audience is, what success looks like, which facts are fixed, what must be avoided, and how the answer should be delivered.
Adding decorative phrases such as “act as a world-class expert” does not repair those gaps. A useful prompt reduces ambiguity. It gives the model enough room to reason while making the important boundaries explicit.
Write a launch email for my app.
Use the G-C-C-F-Q framework
A reliable working prompt can be built from five parts: Goal, Context, Constraints, Format, and Quality check. You do not need every part for every task, but this sequence exposes what is missing before you submit the prompt.
- Goal: describe the outcome, not just the activity.
- Context: provide the audience, product, source material, and relevant situation.
- Constraints: state limits, facts that cannot change, forbidden claims, tone, and length.
- Format: define the structure the answer must follow.
- Quality check: ask the model to verify the answer against explicit criteria.
Write a launch email for a browser-based image compressor aimed at freelance designers. The goal is to earn free-trial clicks. Lead with privacy—files never leave the browser—and avoid unsupported speed claims. Use one subject line, a 70–100 word email, three benefit bullets, and one CTA. Before answering, check that every claim is supported by the supplied product facts.
Give facts, not a fictional job title
Role prompting can help set vocabulary and perspective, but it is not a substitute for evidence. “You are a senior marketer” gives less useful information than the actual customer objection, offer, channel, and conversion goal.
When accuracy matters, paste the source material or identify what sources are acceptable. If information is missing, instruct the model to ask questions or mark uncertainty instead of inventing details.
- Include product facts that may be claimed.
- Separate assumptions from confirmed information.
- Ask for citations only when the model can access reliable sources.
- For code, include versions, interfaces, errors, and existing conventions.
Make quality observable
“Make it better” cannot be tested. “Keep each paragraph under three sentences, include one concrete example, and avoid clichés” can. The more important the task, the more your quality criteria should be visible in the output.
For complex work, ask for a short assumptions section before the answer. This lets you catch a wrong interpretation without demanding hidden reasoning or an unnecessarily long explanation.
After drafting, verify that the answer addresses all four customer objections, contains no invented statistics, and ends with exactly one next action. If a requirement cannot be met, say which input is missing.
Iterate on the failure, not the wording
When an answer disappoints you, diagnose the failure. Was the audience wrong? Was the output too generic? Did the model invent a fact? Did it choose the wrong format? Add the missing boundary rather than endlessly rewriting the whole prompt.
Save prompts that work together with the result and the context in which they worked. A reusable prompt is not a magic sentence; it is a small specification that can be updated as your task changes.
- Generic answer → add audience, examples, and differentiation.
- Incorrect answer → add sources, versions, and an uncertainty rule.
- Wrong tone → provide a short positive example and banned patterns.
- Unusable structure → specify headings, fields, or a machine-readable schema.
Turn the method into a usable prompt
Enter a rough idea and PromptSmith will add structure, constraints, and an output format.
Optimize a prompt free →