AI For Your Business: The CRAFT Framework For Prompts That Actually Work
Most teams get inconsistent results from AI because they’re improvising a new prompt every time. CRAFT (Context, Role, Action, Format, Tone) turns that into a repeatable structure anyone on the team can use, not a skill only one person happens to be good at.
AI Adoption
DOD AI Adoption: how we bring AI into a business safely and consistently.
Ask ten people at a company to prompt an AI tool for the same task and you’ll get ten different qualities of answer. Not because some of them are better writers. It’s because they’re each supplying a different, undocumented mix of context, role, instruction, format and tone, and only some of that mix happens to be enough for the model to work with.
That inconsistency is the single biggest reason AI adoption stalls inside a business. It isn’t a model-quality problem. It’s a structure problem. CRAFT is the fix: five questions, answered in order, every time.
Context
The background the AI has no way of knowing on its own: what's happened so far, what constraints apply, who this is for. Skip it and the model fills the gap with the most generic, statistically average answer it has, which is exactly the output that makes AI feel unreliable for business use.
Role
The perspective the AI should answer from. “Acting as a senior credit risk analyst” and “explain this simply to a new starter” are the same underlying question, but the role changes the vocabulary, the depth, and what gets left unsaid. Most inconsistent output comes from never fixing this in the first place.
Action
One specific, verb-led instruction. Not “help me with the client update” but “draft an email explaining the delay, the reason, and the revised date.” Vague actions get vague output; a single unambiguous task gets something you can actually use.
Format
How the output should be structured, before you get it: word count, bullet list vs. table vs. plain paragraphs, markdown or not. Deciding this after the fact means editing every single response by hand. Deciding it up front means the AI does that work for you.
Tone
The register: a calm, direct client email reads nothing like a Slack message to the team or a formal board memo. Tone is also where brand voice lives. If more than one person on your team is using AI to draft anything external, this is the difference between consistent output and five different “voices” depending on who prompted it.
Worked example
Same situation, one prompt vs. a CRAFT prompt.
Without CRAFT
“Write an email about the project delay.”
With CRAFT
We're two weeks behind on the Acme integration because of a scope change midway through sprint 4. The client sponsor hasn't been told yet.
You are our delivery lead, writing directly to the client sponsor.
Draft an email explaining the delay, the reason for it, and the revised date, and propose one call slot this week to talk it through.
Under 150 words. Three short paragraphs. Plain text, no bullet points, no markdown.
Direct and calm. No corporate hedging language. Sign off as “the team.”
Inconsistent output usually means:
- Context missing, so the model guesses at the situation and hedges everywhere it isn't sure
- No fixed role, so output reads differently depending on how the request happened to be phrased
- A vague action (“help with…”) that gets you an average of every plausible answer, not the one you needed
- Format decided after the fact, so every response gets manually reshaped before it's usable
A CRAFT prompt reliably includes:
- A context line the model can't get wrong or invent
- A role that's explicit, not implied by tone of voice
- One unambiguous, verb-led action
- Format specified before generation, not fixed up after
- Tone that matches who's actually going to read it
Why it matters beyond one person
One person getting good at prompting doesn’t scale. The moment AI use spreads past your early adopters, you need a structure the whole team can follow, one that whoever owns AI governance can actually look at and reason about, rather than trusting that everyone’s individually being careful.
CRAFT is a starting point, not the whole answer. The context you feed a model, the systems it’s allowed to touch, and what happens to the output afterwards all need the same kind of deliberate structure, especially anywhere the output reaches a client, a regulator, or a system of record. That’s the part we spend most of our time on.
Ready to move past one-off prompting?
We help businesses adopt AI with the structure, policy and guardrails that make it safe to rely on, not just a framework for better prompts. Start with a discovery call.