How to Write AI Prompts: 10 Techniques That Actually Work (2026)

What Makes a Good AI Prompt?

A good prompt is like a clear brief to a colleague—it specifies what, how, and why, leaving nothing to interpretation. Most people treat AI like a search engine: they ask a vague question and accept whatever they get. But AI thrives on structure.

The difference between a prompt that gets mediocre results and one that gets exceptional results often comes down to clarity, context, and constraints. A bad prompt wastes your time iterating. A good prompt nails it in one or two tries.

This guide covers 10 techniques that consistently unlock better outputs across ChatGPT, Claude, Gemini, and every major AI model.

Technique 1: The CRISP Framework

CRISP is a mental model that covers everything a good prompt needs:

Context: What's the situation? What's the background?
Role: Who am I, and what expertise should you adopt?
Instructions: What exact task do you need to do?
Style: What tone, format, or approach?
Parameters: What are the constraints (length, audience, depth)?

Example: Weak → Strong

Weak: "Write about remote work"

Strong (CRISP):

Context: I'm a startup founder writing an article for first-time remote workers
Role: You're an experienced tech HR manager who's onboarded 200+ remote employees
Instructions: Write an intro section that covers the top 3 challenges new remote workers face and how to overcome them
Style: Conversational, encouraging, practical (not preachy)
Parameters: 400-500 words, use bullet points for the three challenges, include a real-world example

The second prompt guarantees better output. Here's why: the AI knows exactly who it's writing for, what problem it's solving, and the boundaries.

CRISP Template

Context: [Situation and background]
Role: [Persona or expertise]
Instructions: [Specific task]
Style: [Tone, format, approach]
Parameters: [Constraints: length, audience, depth, format]

You don't need to label everything—just make sure each element is clear.

Technique 2: Few-Shot Examples (Show, Don't Tell)

Telling an AI "be more concise" rarely works. Showing it an example of concise work is incredibly effective.

Example: Summarization

Weak: "Summarize this article in a concise way"

Strong (with examples): "Summarize the following article in 2-3 sentences. Here's the style I want:

Example 1 (article about AI in healthcare): Instead of: 'Artificial intelligence is being integrated into many hospitals across the country to improve patient outcomes through better diagnostic tools and operational efficiency.' Do this: 'AI helps hospitals diagnose diseases faster and run more efficiently.'

Example 2 (article about remote work): Instead of: 'Companies are increasingly allowing their workforce to work from home, which has led to benefits like reduced overhead and challenges like decreased team cohesion.' Do this: 'Remote work cuts costs but can hurt team connection.'

Now summarize this article: [paste article]"

The AI now understands exactly the brevity level and tone you want. Few-shot examples are one of the most underrated techniques in prompting.

Technique 3: Chain-of-Thought (Ask It to Think Step-by-Step)

Chain-of-thought prompting forces the AI to break down reasoning before answering. It dramatically improves accuracy on complex tasks.

Example: Problem-Solving

Weak: "Should I raise prices on my SaaS product?"

Strong (chain-of-thought): "I'm considering raising prices on my SaaS product. Before you give a recommendation, think through these questions step-by-step:

What information do you need to make this decision?
What are the potential upsides and downsides of a price increase?
What risks should I be aware of?
What questions should I ask before deciding?

Once you've thought through those, give me a recommendation."

By asking the AI to show its thinking, you get better reasoning and more defensible conclusions.

Technique 4: System Prompts and Custom Instructions

A system prompt is hidden context that shapes how an AI behaves across all conversations. Most people don't use this, but it's incredibly powerful.

Where to Use System Prompts

ChatGPT: Settings → Custom Instructions → "How would you like ChatGPT to respond?"

Claude (Web): Settings → Custom Instructions

Gemini: Settings → Custom Instructions (rolling out in 2026)

Real Examples

For executives: "You are an executive advisor. I ask you questions about business decisions. Always structure your advice as: 1) what I should do, 2) why, 3) risks, 4) next steps. Be direct and concise—never use more than 200 words."

For developers: "You are a senior software engineer reviewing code. When I paste code, always: 1) identify any bugs or performance issues, 2) suggest the clearest fix, 3) explain why it matters. Use TypeScript and modern best practices."

For writers: "You are a copywriter specializing in SaaS marketing. I'll give you product descriptions, and you'll make them punchy, benefit-focused, and conversational. Avoid buzzwords like 'revolutionary' or 'cutting-edge.'"

Setting this once saves you from repeating context in every single prompt.

Technique 5: Structured Output Formats

Tell the AI exactly how to format its response. This is especially useful for data, comparisons, or lists.

Example: Comparison Table

Weak: "Compare these three AI tools"

Base your comparison on actual 2026 benchmarks, not marketing claims."

By specifying the format upfront, you get structured data you can copy into a spreadsheet or document instantly.

Other Structured Formats

JSON: "Format your response as JSON with keys: tool, pros[], cons[], pricing"
Outline: "Use a numbered outline with main points and 2-3 sub-bullets each"
CSV: "Output as comma-separated values: name, email, company, role"
Markdown: "Use H2 headings for sections, bold for key terms, bullet lists for details"

Technique 6: Constraint-Based Prompting

Constraints force creativity and focus. Instead of "write about X," try "write about X in exactly 100 words" or "using only these three ideas."

Example: Marketing Copy

Weak: "Write a headline for my SaaS product"

Strong: "Write 5 headline options for my SaaS product (TurboAnalytics). Constraints:

Exactly 6-8 words per headline
Focus on the benefit (save time analyzing data), not the feature
Use a verb in the first two options, a question in the next two, a statement in the last one
No exclamation marks
Avoid clichés like 'revolutionary' or 'next-gen'"

This produces five distinct, high-quality options instead of vague alternatives.

Technique 7: Persona-Based Prompting

Assigning a persona makes the AI adopt specific behavior and knowledge.

Persona	Example Prompt
Expert practitioner	"You are a senior data scientist with 15 years of industry experience. Explain how to build a recommender system for a music streaming app."
Journalist	"You are an investigative journalist. What questions would you ask about [topic]?"
Skeptic/critic	"You are a skeptical investor evaluating a business pitch. What are the biggest flaws in this idea?"
Beginner	"You are explaining this to a high school student with no background in the topic. How would you introduce it?"
Storyteller	"You are a screenwriter. Pitch me this product or idea as a compelling story."

The persona triggers different reasoning styles. A skeptic gives you vulnerabilities. A storyteller gives you narrative hooks.

Technique 8: Negative Examples (What NOT to Do)

Show the AI what you don't want. This is sometimes more effective than describing what you do want.

Example: Tone

Instead of: "Write in a professional but warm tone"

Better: "Write in a professional but warm tone. Here's what NOT to do:

Don't be stiff: 'We kindly request your attention to the following matter'
Don't be casual: 'Hey, just wanted to give you a heads-up about this thing'
Don't be passive: 'It has been determined that action is required'

Do this: 'We wanted to make sure you're aware of this change and how it affects you.'"

Negative examples clarify your intent faster than positive ones alone.

Technique 9: Tool-Specific Optimization

Each AI model has strengths. Adapt your prompts accordingly.

ChatGPT

Strength: Conversational, good at explaining concepts
Prompt tip: ChatGPT likes being friendly. Try: "Let's brainstorm together" instead of "Generate ideas"
Weak area: Reasoning on novel problems
Hack: Use examples and ask it to identify the pattern before solving your problem

Claude (Anthropic)

Strength: Nuanced reasoning, long documents (200k token context)
Prompt tip: Claude responds well to explicit constraints. Be very specific about edge cases
Weak area: Rapid-fire back-and-forth conversations
Hack: Paste entire docs and ask Claude to analyze or edit them—it handles long context better

Gemini (Google)

Strength: Real-time web search, Google integration, image understanding
Prompt tip: Gemini appreciates specificity about sources. Try: "Answer based on current 2026 data"
Weak area: Complex coding or math
Hack: Ask it to search the web for context before answering complex questions

Model	Best For	Avoid
ChatGPT	Quick answers, brainstorming, explanations	Long documents (context limit), novel reasoning
Claude	Deep analysis, long content, reasoning, coding	Real-time data, speed (slower processing)
Gemini	Web research, Google integration, current data	Specialized coding tasks, reasoning chains

Technique 10: Iterative Refinement With Feedback

Most people write one prompt and accept the result. Better approach: treat it as a conversation.

Example Iteration

Prompt 1: "Write a job description for a marketing manager" Result: Generic, templated

Prompt 2: "That's too generic. Make it more specific to our company culture (startup, fast-paced, data-driven). What details are missing?" Result: AI identifies what's vague and asks clarifying questions

Prompt 3: "We're a series B SaaS startup, 50 people, focus on developer tools. This person owns demand gen. They'll report to the VP of Growth. Make it reflect our casual but professional tone." Result: Much better—specific, cultural, actionable

The iteration loop gets you from 60% to 95% in 2-3 refinements.

Common Mistakes That Kill Prompt Quality

Being too vague: "Tell me about productivity" vs. "How can a solo founder manage task switching and maintain focus while handling sales, support, and product?"
Forgetting context: The AI doesn't know your industry, audience, or constraints unless you say so.
Asking for too much at once: Break complex tasks into steps instead of asking for a 2000-word essay plus a table plus code examples.
Ignoring the output format: Always specify how you want the answer structured (list, table, essay, JSON, etc.)
Not leveraging examples: Showing one good example beats describing quality in 100 words.
Using vague adjectives: "Make it better," "more concise," "professional" are meaningless without context. Use "cut it to 150 words" or "explain it to a 5th grader."
Treating it like a search engine: Asking "what is X?" gets surface-level answers. Ask "explain X in the context of my situation" for relevant output.

Pro Tips You Rarely See Shared

Use "Let me show you what I mean": Pasting context, examples, or draft work often produces better results than describing it
Ask the AI to find the gap: "What information am I missing to make this decision?" gets the AI to identify blind spots
Request iteration formats: "Generate 3 versions: one for executives, one for the team, one for customers" produces multiple angles efficiently
Use "What if" prompts for brainstorming: "What if we removed [constraint]? How would that change the strategy?"
Ask for a rubric: Before getting an output, ask "What would an excellent [output] include?" to align expectations
Combine prompt techniques: Use CRISP + few-shot examples + chain-of-thought together for complex tasks

The Prompt Evolution Framework

Think of prompting as iterating toward clarity:

First draft: Simple request
Add context: What's the background and goal?
Add constraints: Length, format, audience, tone
Add examples: Show what good looks like
Ask for reasoning: Chain-of-thought before the answer

Each step compounds. A bare request becomes a machine-readable specification.