The CRAFT Prompt Formula: A Practical Framework for Better AI Outputs

You've probably experienced this: you feel like you've explained things clearly, but the AI's response is completely off from what you wanted. You ask for a "summary" and get a detailed analysis. You ask for "brief" and get 2000 words. You ask for "suggestions" and get a lecture.

The problem probably isn't the AI. It's the prompt.

After using AI tools for over two years and going through countless rounds of trial and error, I've settled on a framework I find genuinely useful -- CRAFT. It's not some deep academic theory. It's just a way of breaking down "how to explain things to an AI" into a few consistent components that ensure you're giving the AI all the information it needs to produce exactly what you want.

What Is CRAFT

Five letters, five elements of a well-structured prompt:

• Character (Role): Who should the AI be
• Request (Task): What should the AI do
• Action (Steps): How should the AI do it
• Format (Structure): What format should the output be in
• Target (Audience): Who is this output for

You don't need all five every time. A simple task might only need R and F. Complex tasks benefit from the full set. But knowing these five dimensions exist helps you not leave out critical information when writing prompts. Think of it as a checklist -- before hitting enter, quickly run through each letter and consider whether adding that element would improve your result.

C -- Character: Tell the AI Who It Is

AI has broad knowledge, but it doesn't know which part to use for your request. Giving it a role helps it narrow down its response to the most relevant style, tone, and level of expertise.

Basic version:

You are an experienced internet product manager who writes good PRD documents.

More detailed version:

You are a senior copywriter who specializes in tech-focused articles.
Your style is: conversational, insightful, not stuffy.
You're good at explaining technical concepts in plain language without dumbing them down.
Your readers are internet professionals aged 25-35 who already understand basic tech concepts.

The more specific the role, the more targeted the AI's response. "You are an expert" is too vague -- expert in what? With what perspective? "You are a growth specialist with 8 years of experience focused on B2B SaaS company user acquisition" is an effective role definition because it tells the AI exactly which knowledge base to draw from and what lens to apply.

R -- Request: Be Clear About What You Want

This is where most people's prompts fall short. They assume the AI can read their mind about what constitutes a good response.

The most common approach (and why it doesn't work):

Help me write a proposal
Help me revise this article
Give me some suggestions

These kinds of descriptions give the AI nothing to work with. It doesn't know who the proposal is for, what scenario it's meant for, what outcome you're after, or what "revision" or "suggestions" even means in your context.

A better approach: three elements of task description

1. Verb: What specific action to take (write, analyze, revise, summarize, compare...)
2. Deliverable: What the final output should be (a document, a table, a paragraph, a list...)
3. Goal: What effect you want to achieve

Improved version:

Your task is to help me write a proposal for introducing AI tools to the management team.
The deliverable is a roughly 1,000-word document covering: background, overview of similar tools,
expected outcomes, and cost estimate.
The goal is to help management understand why we should introduce this tool and approve the budget.
The audience has technical literacy but may not be familiar with AI-specific terminology.

This version gives the AI clear guardrails. It knows exactly what to produce (a proposal), what sections to include (background, overview, outcomes, costs), how long to make it (~1000 words), and what success looks like (getting budget approval).

A -- Action: Tell the AI How to Proceed

For complex tasks, asking the AI for a final result directly often leads to missed steps or muddled logic. Giving it a step-by-step process works much better.

For simpler tasks:

Please follow these steps:
1. First, analyze what the core problem is with this requirement
2. List 3 feasible solutions
3. Do a pros/cons analysis for each solution
4. Give your recommendation and reasoning
5. Suggest next steps for implementing the recommendation

For complex tasks (phased approach):

Please follow this workflow, completing each phase before moving to the next:

[Phase 1: Information Gathering]
Ask me 3 key questions to make sure I've provided enough information.
Don't proceed until you have these answers.

[Phase 2: Framework Setup]
Based on my answers, output an outline first. Include the main sections,
key points under each, and estimated word count.
I'll confirm before you continue.

[Phase 3: Writing]
Follow the confirmed outline and write section by section.
After each section, briefly explain its main point (2-3 sentences).

The benefit of phased execution: you can check whether the direction is correct at each stage, rather than discovering at the end that the AI went off track. This iterative approach produces much higher quality results for complex tasks and saves time compared to reworking a completely wrong output.

F -- Format: Make the Output Usable

Without format requirements, AI tends to give you a wall of text that you then have to reformat yourself. Being specific about structure saves you cleanup time.

Common format instructions:

Output in Markdown format
Use tables to compare different options
Keep each point to no more than two lines for readability
Bold key content for emphasis
Use emoji to mark priority levels: (🔴 high 🟡 medium 🟢 low)
Use numbered steps for procedures

A more complete format definition:

Output format requirements:
# Title (concise, under 10 words)
## I. Core Takeaway (one-sentence summary of the key finding)
## II. Detailed Analysis (bulleted, no more than 3 lines per point)
## III. Action Items (in table format: Action | Priority | Estimated Time)

Requirements:
- Concise language, no repetition or filler
- Specific data instead of vague words like "many" or "a lot"
- Every recommendation should be immediately actionable
- Include numerical estimates where possible

Clear format instructions are especially important when you plan to share the output with others. Having AI produce the output in a format you can use directly (email, report, presentation) saves significant post-processing time.

T -- Target: Who Is This For

The same content reads completely differently depending on whether it's for a CEO, a colleague, or a general audience. The tone, level of detail, focus, and vocabulary all change.

For a CEO:

This report is for the CEO. She's time-constrained and only cares about conclusions and numbers.
Don't include process details -- go straight to conclusions and recommendations.
Format: Executive summary at the top (3 bullet points), details below only if she asks.

For a technical team:

The readers are senior software engineers who know the codebase well.
You can include technical details and API references without explanation.
Focus on architectural decisions and implementation details.
Use terms they're familiar with -- don't over-explain basics.

For non-technical readers:

The reader has no technical background at all.
Avoid jargon completely. If you must use a technical term, explain it in simple language.
Use analogies and everyday examples to illustrate concepts.
Focus on what the technology does, not how it works internally.

Specifying the target audience helps the AI calibrate its language complexity, depth, and perspective appropriately.

A Complete Example

Putting all five elements together, a full CRAFT prompt looks like this:

[C -- Character]
You are a digital marketing consultant with 10 years of experience,
specializing in social media strategy for small and medium businesses
in the Chinese market. You have managed accounts across Douyin,
Xiaohongshu, WeChat, and Bilibili.

[R -- Task]
Help me create a cold-start plan for a Xiaohongshu (Little Red Book) account.
My account focuses on "workplace productivity tool reviews," targeting
young professionals aged 22-30 in Tier 1 Chinese cities.
The deliverable is an actionable 30-day plan for growing from 0 to 1000 followers.
Budget: under 2000 RMB for paid promotion.

[A -- Steps]
1. First, analyze the competitive landscape -- what types of accounts are saturated,
   where there's still opportunity in the productivity niche
2. Give account positioning advice -- where's the differentiation opportunity
3. Create a 30-day content calendar (what themes to post each week)
4. List 5 content directions most likely to go viral in this niche
5. Suggest a paid promotion strategy within the budget

[F -- Format]
Structure the output as:
# Account Positioning
## Competitive Analysis
## 30-Day Content Plan (table: Date | Theme | Format | Expected Goal)
## Viral Content Directions (each with a specific title example)

[T -- Audience]
This plan is for me -- a solo operator with no team and limited editing skills.
So keep recommendations practical -- nothing that requires multiple people,
professional equipment, or advanced video editing to execute.
Most content should be postable from a smartphone within 15-30 minutes.

A prompt like this will produce dramatically better results than "help me grow my Xiaohongshu account" because it gives the AI all the context, constraints, and guidance it needs.

A Few Practical Tips

1. Simple tasks don't need the full CRAFT treatment. Asking "what's the weather today" doesn't need a role definition, audience specification, or phased approach. CRAFT is a framework for complex tasks, not something to force onto every sentence. Use your judgment about which elements add value for a given prompt.

2. Start with R, then add the rest. Many people get stuck on how to write the character. My advice: get the task (R) clear first -- that's the core that determines output quality. Once you have the task defined, the other elements (C, A, F, T) are refinements that further improve quality.

3. Iteration beats perfection. After your first prompt, look at the AI's output and adjust based on what you see. If the output is too long, add "no more than 3 lines per section." If it's too general, add "give 3 specific examples with real brand names." If the tone is off, add "use a more conversational tone." Tweak based on what you actually get back rather than trying to specify everything in advance.

4. Save prompts that work. If you find a prompt structure that consistently gives good results, save it as a template in a dedicated document or note-taking app. Next time you have a similar task, just change the specific details instead of starting from scratch. Over time, you'll build a personal library of reliable prompt templates.

One Last Thing

CRAFT isn't a magic bullet. It won't guarantee perfect answers every time, but it significantly increases the odds of getting a useful result on the first try. In my experience, using CRAFT consistently has cut my "revision cycles" (the number of follow-up prompts needed to get a usable result) roughly in half.

Writing prompts is really about training yourself to express your ideas clearly. That skill isn't just useful for AI -- it helps with communicating with people too. Clear thinking and clear writing go hand in hand, and the practice of writing good prompts sharpens both.