A great AI prompt is not a magic phrase. It is a clear brief. The difference between a reply you throw away and one you actually use almost always comes down to how much of the thinking you did before you hit send. This guide breaks down what makes a prompt strong, shows a weak-to-strong rewrite, and gives you a template you can copy today. If you want to skip the theory, try the generator and edit what it produces.
What separates a great prompt from a weak one
Weak prompts assume the model can read your mind. "Write a blog post about coffee" gives the model a thousand valid directions, so it picks the blandest average of all of them. Strong prompts remove the guesswork. They state who is writing, for whom, what the output should contain, and how to format it.
Three habits do most of the work:
- Be specific. Replace "good" and "engaging" with concrete criteria: length, reading level, the point the reader should walk away with.
- Give context. The model knows nothing about your product, your reader, or last week's meeting unless you tell it.
- Constrain the output. Say what to include, what to leave out, and the exact shape you want back (a table, five bullets, JSON, a 200-word paragraph).
A prompt that does these three things will beat a clever one-liner almost every time, across ChatGPT, Claude, and Gemini alike.
The anatomy of a strong prompt
You rarely need all of these at once, but the best prompts are assembled from the same parts. Think of them as ingredients you add as the task gets harder.
1. Role or persona
Tell the model who it should be. "You are a technical recruiter with ten years in fintech" sets vocabulary, priorities, and tone in one line. A role is shorthand for a whole set of assumptions, so it saves you from spelling each one out.
2. Objective
State the single job in plain language. "Write a rejection email that keeps the candidate warm for future roles." One clear objective beats three vague ones. If you have several goals, rank them.
3. Context and audience
Give the facts the model cannot infer: who reads this, what they already know, the product details, the constraints of the situation. "The reader is a non-technical founder evaluating whether to hire in-house or outsource" changes every sentence that follows.
4. Step-by-step instructions
For anything with moving parts, number the steps. "First summarize the transcript in three bullets, then list action items with owners, then flag anything left unresolved." Ordered steps stop the model from skipping the boring middle.
5. Output format
Describe the exact shape you want. A markdown table with three named columns. A JSON object with these keys. Exactly five bullets, each under fifteen words. Format instructions are the single cheapest way to make a reply usable without editing.
6. Constraints
Set the guardrails: word count, tone, reading level, what to avoid, which sources to rely on. "Under 120 words, no jargon, do not invent statistics" prevents the three most common failures in one line.
7. A short example
When tone or structure matters, show one. "Match the style of this example: [paste]." One good example teaches faster than a paragraph of adjectives. This is called few-shot prompting, and it is covered further in prompt engineering basics.
Weak vs strong: a before and after
Here is the same request written twice. The weak version is what most people type. The strong version is what gets a usable answer on the first try.
Weak: Write a product description for my running shoes. The model has no idea what the shoe is, who buys it, how long the copy should be, or where it will appear. The result is generic filler that could describe any shoe on earth.
Strong: You are a copywriter for a direct-to-consumer running brand. Write a product description for the Trailhead 2, a lightweight trail shoe for weekend runners who want grip on wet terrain. Audience: recreational runners aged 30 to 50 who care about comfort over speed. Include: one benefit-led opening line, three short bullets on grip, cushioning, and weight, and a closing call to action. Constraints: under 90 words, warm and plain, no superlatives like "best" or "revolutionary," do not invent specs I have not given you. The second prompt names the role, the product, the reader, the exact structure, the length, and the tone, and it forbids the two things models tend to get wrong: hype and invented facts. Same task, completely different result.
Common mistakes
Most disappointing outputs trace back to one of these.
- Too vague. "Make it better" gives the model nothing to act on. Say what "better" means: shorter, friendlier, more concrete, fewer buzzwords.
- No format. If you do not ask for a structure, you get a wall of prose you then have to reshape by hand.
- No constraints. Without a word count or a tone, the model defaults to long and generic. Boundaries improve quality.
- Overloading one prompt. Asking for a strategy, the copy, the SEO tags, and a social plan in a single message spreads the model thin. Split big jobs into steps and feed the output of one into the next.
- Assuming shared context. The model did not sit in your meeting. Paste the relevant facts instead of referring to them.
How to iterate
Treat the first answer as a draft, not a verdict. The fastest way to improve a result is to keep the prompt and change one thing at a time.
- Diagnose. Name what is wrong: too long, wrong tone, missed the point, invented a detail.
- Adjust one variable. Add a constraint, tighten the objective, or supply a missing fact. Changing five things at once tells you nothing about what helped.
- Ask the model to critique itself. "List three ways this could be sharper, then rewrite it applying them" often beats another round of you rewriting the prompt.
- Save what works. When a prompt lands, keep it. A small library of proven prompts is worth more than any single clever trick.
How text, image, code, and video prompts differ
The core idea, be specific and give context, holds everywhere, but each medium rewards a different emphasis.
| Type | What to prioritize |
|---|---|
| Text | Role, audience, format, and tone. Constraints on length and reading level matter most. |
| Image | Subject, style, composition, lighting, and camera or medium. Word order and specific nouns carry weight. See Midjourney prompts. |
| Code | Language and version, inputs and outputs, edge cases, and whether you want an explanation. State the runtime and any constraints. |
| Video | Shot description, motion, duration, camera movement, and pacing. Describe the action across time, not just a still frame. |
For chat models specifically, our guide to ChatGPT prompts goes deeper on roles and follow-ups, and AI prompt examples collects ready-made prompts you can adapt.
A copy-paste template
Fill in the brackets, delete any line you do not need, and paste it into your model of choice.
Role: You are a [role] with expertise in [domain].
Objective: [the single main task].
Context: [who the reader is, what they know, product or situation details].
Instructions:
1. [first step]
2. [second step]
3. [third step]
Output format: [table / bullets / JSON / word count and structure].
Constraints: [tone, length, reading level, what to avoid, do not invent facts].
Example to match (optional): [paste a sample].Start with the role, objective, and output format; add context and constraints as the task gets harder. When a result misses, change one line and run it again. That loop, more than any single phrase, is what turns a weak prompt into a great one. Ready to build one now? Try the generator.