# The Prompting Craft

By [DYLIT Chronicles](https://dylit.info/user/dylitmediabuzz)

[Everything AI - beyond the hype](https://dylit.info/pr/everything-ai-beyond-the-hype/6a9efac02e92664f4d50cf9d) > [The Prompting Craft](https://dylit.info/ch/the-prompting-craft/6ab358fa99638e119f592c7d)

C-R-A-F-T: Five Parts of an effective AI Prompt Most bad AI output traces back to a prompt that was missing something the model had no way to supply on its own. Not missing a clever phrase. Missing information. The reader it was written for. The constraint it had to respect. The format it needed to arrive in. People type a nine-word instruction, receive nine words' worth of usefulness, and conclude the tool is overrated. C-R-A-F-T is a checklist for the sixty seconds before you hit send. Context, Role, Action, Format, Test. You will not need all five every time. For anything that matters, running through them roughly doubles the quality of what comes back, and it costs less time than rewriting the answer. C: Context The situation, the background, the facts that bear on this. This is the part people skip most and it is the part that matters most, because the model knows nothing about your world. It has never seen your product, your team, your customers, or last quarter's results. Every relevant detail you leave out is a gap, and gaps do not come back as questions. They come back as confident invention. Weak: "Write an update about the project delay." Better: "Our integration is slipping two weeks because a vendor dependency landed late. This is a long-standing enterprise client and it is our first delay with them." Same task. The second version has something to work with. A useful instinct: give more context than feels necessary. It is nearly impossible to over-explain background to these systems. It is very easy to under-explain it. R: Role The perspective the model should take. This works because the model learned patterns from an enormous range of writing, and naming a role steers which patterns it draws on. Text written by a financial controller genuinely does differ from text written by a marketer, and the model absorbed that difference. The trick is specificity. "Act as a finance expert" is nearly useless, because it names a category rather than a person with a viewpoint. "Take the role of a controller at a mid-size manufacturer who has been burned before by optimistic revenue forecasts" gives the model something to actually do. Roles are most valuable when you want a particular kind of scrutiny. A skeptical buyer reading your pricing page will find different problems than a generic request to review it. A: Action The exact task, stated as a verb. Vague asks produce vague output, and this is where most prompts quietly fail. "Look at this" and "help me with this" do not specify an operation. Summarise, rank, compare, rewrite, extract, critique, translate, categorise: each of these is a different job with a different result. Be specific about scope too. "Identify the three most likely causes" produces something usable. "Analyse the situation" produces an essay. If you want more than one thing done, do them in separate messages. Bundled requests dilute every part of the answer. The model spreads its effort across all of them and does none particularly well. F: Format Structure, length, tone. The model has no default sense of what you will do with the output. A table, five bullets, two paragraphs, a formal letter, a Slack message: all are equally available, and if you do not choose, you get whatever the request's phrasing suggested, which is usually an essay. State the shape. "A table with columns for driver, evidence, and confirming test. Under 200 words." Naming a word limit does more work than people expect, because it forces prioritisation. A constrained answer has to lead with what matters. T: Test The constraints and success criteria the answer has to satisfy. This is the least used and, for anything with real stakes, the most valuable. You are telling the model what would make the output wrong. "Use only the information in the notes below. Do not guess. If something is not covered, write 'not stated' rather than filling it in." "Flag anything you are uncertain about rather than smoothing over it." "Do not offer legal advice. Where an issue needs a lawyer, say so." These instructions genuinely change the output. They will not eliminate errors, and you still have to check, but you are substantially narrowing the space in which the model invents things. All five, on one request Take a real example. Renewal rate dropped from 91 percent to 84 percent, and you have the exit survey notes. Context: "Our Q3 renewal rate fell from 91% to 84%. The exit survey notes are included below." Role:  "Take on the role of a customer success director." Action : "Identify the three most probable drivers, and specify what evidence would confirm each one." Format : "A table with columns for driver, supporting evidence, and confirming test. Keep it under 200 words." Test : "Rely only on the provided notes. Do not guess, and flag anything the notes do not back up." Roughly ninety words of setup. What comes back is an analysis you can take into a meeting. Compare it to "why did our churn go up," which produces a competent general essay about customer churn that would apply equally to any company on earth. When to skip parts Role is often unnecessary for mechanical tasks. Reformatting a list does not need a persona. Test matters most when the model might fill gaps, which is to say whenever it is working with your data or your documents. Context is the one to skip least. If you only have time for one element, give it the background. The thing nobody mentions The real benefit of running through C-R-A-F-T is not what it does to the model. It is what it does to you. Half the time, working through the five parts reveals that you had not decided what you wanted. You did not know who the document was for, or what decision it needed to support, or how long it should be. You were going to find out by writing it. Sixty seconds of structure surfaces that before you have generated three pages in the wrong direction. Which is useful on its own, whether or not you ever open the tool.