Five letters.One repeatable system.
CRAFT is the proprietary framework we teach in every cohort. It turns prompting from art into craft — a reusable playbook you can execute under pressure, on any model, for any task. Five slots, every time: Context, Role, Action, Format, Tone.
A framework is only useful if it holds up under pressure.
CRAFT has been stress-tested across thousands of prompts in dozens of industries. Three properties are what make it durable.
Repeatable
The same five slots, every time. You never start from a blank page — you start from a scaffold you already trust.
Portable
Works in Claude, ChatGPT, Gemini, Grok, and every future model. The models change; the craft doesn't.
Teachable
Your team can execute it without you. CRAFT is how you stop being the prompt bottleneck inside your own company.
Context
Give the model the world it needs.
What it is
Context is everything the model needs to know before it starts thinking: who the answer is for, what situation you're in, what data or documents it should ground in, what's been tried before, and any constraints it has to respect. If you had to hand this task to a smart new hire, everything you'd tell them in the first five minutes is context.
Why it matters
In every cohort, when we A/B-test identical prompts with and without rich context, the context-loaded prompt wins by a wide margin — often 3-4× in subjective rater studies and even more in task-specific evals. Context is the single highest-leverage lever in a prompt, and it's also the one most people skip because they assume the model already 'knows'.
Weak prompt
“Write a LinkedIn post about leadership.”
CRAFT-formatted — Context slot
“Context: I run a 12-person real estate fund focused on manufactured housing communities. My audience on LinkedIn is 60% LPs (RIAs, family offices, small institutions), 30% fellow GPs, and 10% operators at my portfolio companies. A peer just posted a generic 'servant leadership' take that got 40K impressions — I want to add to that conversation rather than echo it. My angle: leadership in distributed ops teams you only see quarterly.”
Where people go wrong
- Dumping raw context without a label — the model doesn't know where context ends and the task begins. Always use a 'Context:' header.
- Leaving out the audience. 'Who is this for' is half the context.
- Assuming the model remembers prior turns. For important facts, restate them every prompt.
Role
Pin the model to a persona and level of expertise.
What it is
Role tells the model who to be and for how long. It compresses a huge amount of instruction into a single sentence. A senior editor at The Economist, a CMBS attorney with 500 deals under their belt, a VC partner known for skeptical first principles thinking — each of these roles pulls a very different style of reasoning out of the same base model.
Why it matters
Role works because LLMs were trained on enormous amounts of domain-specific text and they carry style and reasoning patterns with them. Invoking a role activates the distribution of writing that matches it. 'Be helpful' is a null signal; 'you are a senior M&A attorney who has negotiated 500+ purchase agreements on behalf of buyers' is a loud one.
Weak prompt
“Check this contract for issues.”
CRAFT-formatted — Role slot
“Role: You are a commercial real estate attorney with 15+ years negotiating CMBS loan agreements on behalf of borrowers. You've seen hundreds of these docs and you have a reputation for flagging unusual covenants that lenders try to sneak in. You have a low tolerance for boilerplate and a high tolerance for reading the fine print.”
Where people go wrong
- Flattery roles: 'you are a genius.' These give the model nothing to work with. Specificity beats superlatives.
- Roles that contradict the task: asking 'a comedian' to write a SOC-2 audit report.
- Forgetting to pair the role with a seniority or years of experience — those anchors change the register meaningfully.
Action
Name the verb clearly, and only one at a time.
What it is
Action is the verb. Draft, critique, compare, rank, extract, rewrite, summarize, translate, score. It's the most concrete part of a prompt and also, weirdly, the part most people get wrong by being vague. 'Look at this', 'tell me what you think', 'give me feedback' — these are not actions, they're invitations for the model to do whatever it wants.
Why it matters
Every coherent prompt has a verb at its center. When that verb is ambiguous, the model substitutes its own interpretation and you get drift. When the verb is precise, the output is too. 'Rank these five offers by NPV using a 12% discount rate' produces a table; 'what do you think of these offers' produces a ramble.
Weak prompt
“What do you think about this offer?”
CRAFT-formatted — Action slot
“Action: Compare the three LOIs on these criteria, weighted as follows — price per pad (60%), earnout exposure (25%), closing certainty (15%). Produce a weighted score for each LOI (0–100 scale) and rank them. Then, in a second pass, identify the single weakest term in the winning LOI and draft suggested redline language for it.”
Where people go wrong
- Combining three actions in one sentence without labels. Break them into Step 1, Step 2, Step 3.
- Using 'analyze' — too generic. Swap in 'compare', 'rank', 'score', or 'extract'.
- Hidden actions in the context. If you want a critique, say 'critique', don't bury it under 'tell me about it'.
Format
Structure is a feature, not an afterthought.
What it is
Format is the shape of the answer. A markdown table, a bulleted list, a JSON object with a specific schema, a three-paragraph memo, a 280-character tweet. Picking the right container for the content is one of the easiest ways to turn a blob of text into a decision-grade artifact.
Why it matters
Format doesn't just make output easier to read — it constrains the model's generation process itself. Asking for a table with specific columns forces the model to surface a fact per cell, which meaningfully reduces hallucination and surfaces missing data. Asking for a JSON schema makes the output machine-parsable in one step.
Weak prompt
“Explain the pros and cons of the three pricing tiers.”
CRAFT-formatted — Format slot
“Format: A markdown table with Tier (Starter / Pro / Team) as rows, and these columns: Price / Month, Core Features, Best For, Ideal Customer Profile, Main Risk. After the table, write a 2-paragraph recommendation for which tier best fits a 10-person creative agency, and end with a 1-line TL;DR.”
Where people go wrong
- No format spec at all — the model defaults to wall-of-text.
- Format contradicts action. 'Give me a 3-page memo summarized in one tweet' is incoherent.
- Over-specifying format for creative tasks. When voice matters, leave room for prose.
Tone
Voice and register do the invisible work.
What it is
Tone sets the voice and register of the output. Confident, academic, warm, irreverent, skeptical. Patagonia's brand voice. Your own personal Slack style. The difference between a cold email and an investor memo about the exact same deal is almost entirely tone.
Why it matters
Two drafts can be factually identical and read like they came from different humans. Tone is what turns a correct draft into a credible one. Models default to a bland, corporate-safe tone unless you steer them off it — which is why so much AI-written content pattern-matches to 'obviously AI'.
Weak prompt
“Write a cold email to this prospect.”
CRAFT-formatted — Tone slot
“Tone: Confident but not pushy. Writer is a senior operator, not a junior SDR. Short sentences. No 'Hope you're having a great week!' opener. Zero exclamation points. Assume the reader is time-starved, reads 50 cold emails a day, and is allergic to LinkedIn-speak. Aim for the voice of a 40-year-old partner who has done this 1,000 times.”
Where people go wrong
- Asking for 'professional' — too vague. Name a specific writer, publication, or brand voice to imitate.
- Forgetting what the reader has already seen. 'Confident' to a skeptic reads as 'arrogant'.
- Letting tone drift mid-output. For long pieces, give tone rules at top and again at the transition points.
All five slots. One prompt.
Here's what a full CRAFT prompt looks like in the wild — the kind we build together in Week 4. Notice that every slot is labeled, and there's no ambiguity anywhere.
Three weeks to makeCRAFT muscle memory.
CRAFT is Part 02 of the Level Up With AI cohort. Three weeks of live sessions, worked examples across 12 knowledge-worker tasks, and a template library you keep forever. By Week 6, you'll be running CRAFT without thinking about it.