CRAFT Prompting Framework
Structure AI requests with context, role, action, format, and target
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 99%
CRAFT treats prompting as the act of briefing a highly capable but inexperienced colleague. The user first supplies context about the situation and relevant source material, then assigns the model a role that frames its perspective. Next, the user defines the action to perform, specifies the format and tone of the result, and identifies the target audience. These inputs constrain the model while giving it enough information to produce something useful. When the first result misses the mark, the user diagnoses which CRAFT component was incomplete rather than merely asking the model to try again. Software vendors can eventually automate parts of the framework by supplying customer, channel, and workflow context behind the scenes.
Origin
Nicholas Holland shared CRAFT as his preferred model for working with an AI that behaves like a smart white-collar intern. Extracted from Marketing Against The Grain.
Core principles
- 01Treat AI like a capable intern that still needs direction
- 02Supply relevant background before requesting an output
- 03Define the role and desired action explicitly
- 04Specify the output format rather than assuming it
- 05Adapt every result to its intended audience
How to run it
- 1
Load the context
Explain the situation, objective, constraints, and relevant history. Include source material that the model cannot reliably infer.
Pro tip Supply representative examples of previous work when voice or style matters.
Watch out Uploading public material may add little differentiation if the base model already knows it.
- 2
Assign the role
Tell the model what function or perspective it should adopt, such as a BDR, social strategist, or editor.
Pro tip Choose a role with responsibilities that match the requested result.
Watch out A broad role such as "marketer" may leave too many decisions unresolved.
- 3
Define the action
State precisely what the model must produce or accomplish. Include the business objective behind the task when it affects the answer.
Pro tip Use an observable action such as draft, rank, compare, or diagnose.
Watch out Do not confuse background information with the actual instruction.
- 4
Specify the format
Describe the structure, channel, length, tone, and stylistic conventions required for the final output.
Pro tip Model the format on the destination, because LinkedIn, email, and X require different treatments.
Watch out Without format constraints, the model may default to bland internet-average content.
- 5
Identify the target
Describe who will receive the output and what matters to them. Include role, sophistication, pains, and relationship context where relevant.
Pro tip Distinguish between peers, executives, customers, and friends even when the underlying message is similar.
Watch out Personalization without useful recipient context can feel automated or inauthentic.
- 6
Review and refine
Evaluate the artifact against every CRAFT component and add whatever information was missing. Continue until the result is ready for real use.
Pro tip Correct the brief before rewriting individual sentences by hand.
Watch out Do not blame the model for errors caused by omitted constraints or context.
In the wild
A BDR gives the model CRM context, assigns it the BDR role, requests a short meeting email, specifies a punchy company voice, and describes a busy CMO who hosts a podcast. The model uses those constraints to draft an audience-aware message rather than a generic sales email.
→ The team receives a concise outreach draft grounded in the account and recipient.
A social manager supplies past posts and campaign context, asks the model to act as a LinkedIn strategist, requests a launch post, defines the desired tone and structure, and describes the target customer. The manager then reviews whether each CRAFT element appeared in the draft.
→ The output fits the channel, brand voice, and intended audience more closely.
Common mistakes
Providing examples without a format
Source content alone does not tell the model how the final artifact should be structured, voiced, or delivered.
Leaving the audience implicit
The same message should change when addressed to a customer, peer, executive, or friend.
Treating the first draft as final
A weak result often reveals missing context or constraints that should be added to the brief.
Is it for you?
Best for
It is best for marketers using generative AI to create content, strategies, outreach, or other production-ready artifacts.
Not ideal for
It is not ideal for simple factual questions that require little context or audience adaptation.
From the transcript
“craft has become like this like my favorite of mine model of working with now this white collar intern”
“context you basically tell it what you're trying to get done give it a bunch of background”
“actions you tell it what you're trying to get done out of this”
From the episode
Tools & Strategies You MUST Use To Survive The 2024 AI Revolution