A Practical and Memorable Framework to Generate Consistent, High-Quality Outputs from AI (LLMs)
Most folks use AI in sub-optimal ways, including some of the tech professionals. I have been in Technical Product Management for years and even I make basic mistakes every now and then. The issue is that most people are confident (erroneously) of their own ability to communicate or to describe a problem and then they just wing it. If you are doing that, you are leaving a lot on the table. What we need is a prompt engineering framework that works for us tech professionals which is memorable, and can deliver with more than 90% of what we need regardless of the context or domain. Here is my attempt at such a framework – CRAFT.
CRAFT Framework – One Acronym, Five Pillars, No BS.
The Framework at a glance:
| Letter | Pillar | What it covers | Priority |
|:—:|—-|—-|—-|
| C | Context & Role | What’s the Situation? + Who is the AI? | Core |
| R | Request & Reason | What is the Task and Goal (Why)? | Core |
| A | Augment | Examples + Chain-of-thought reasoning | Core |
| F | Format & Fences | Output structure + Data Delimiters | Advanced |
| T | Trust Boundaries | Constraints + Guardrails + What NOT to do | Advanced |
n C, R and A are the core pillars – DO NOT miss these.
F & T are the ones that differentiates great prompts from good prompts. It’s up to you. 🙂
C – Context and Role
Define the world that the AI is operating in and the persona that the AI must assume. This one aspect probably has the greatest impact on the quality of output that you get from an LLM.
Include: expertise level, industry, company stage, goals, and constraints.
Bad example: “You are a PM working for a tech company”
Good Example: "You are a senior B2B SaaS PM with 10 years of experience in enterprise software,
specializing in PLG strategies. Our company is post-Series B, 80 employees,
expanding into the EU market, targeting mid-market HR teams."
You can see that each work in the example above will enable the LLM to narrow down on the knowledge, frameworks and skills needed to be specific with the user’s needs.
R – Request
Tell the AI exactly what task needs to be done and why (Goal) does this task matter? Make sure that you start with an action verb.
Bad Example: Need details on our Q3 metrics
Good Example: "Analyze Q3 metrics and identify the top 3 growth drivers, framed by ARR impact,
for a board presentation."
A – Augment
Give a high-level instruction to the AI on how to think through the task in hand. Here are the two main components of augmentation:
Chain-of-thought reasoning: Ask the AI to reason in a step-by-step manner. This reduces hallucinations:
Bad Example: Do it meticulously and precisely.
Good example: Think step by step. First identify key variables, then analyze relationships,
then form a recommendation."
Examples (Few-shot examples): Give the AI 2-3 examples of Input/output matches the way you want it. Product Managers can think of this as an acceptance criteria.
F – Format and Fences
Specify exactly how the oujtput must be structured.
Bad example: Give the output in JSON
Good example: Return as JSON:
{"summary": "...",
"key_findings": [...],
"recommended_action": "..."}
T – Trust Boundaries
This part is what makes a prompt production-grade. Tell the AI what it cannot do and what to do when the AI is uncertain.
Here are some examples
Constraints:
Bad example: Keep it short and focussed.
Good example: Keep under 250 words. Focus only on Q3 data. No competitor analysis.
Negative Instructions:
Bad example: DO NOT create subpar content or hallucinate.
Good example: Do NOT fabricate statistics. Do NOT provide legal advice
Guardrails:
Bad example: Give a safe answer for questionable/uncertain scenarios.
Good example: If you're unsure, say: 'I don't have enough information to answer this reliably.
Beyond the framework
Iterate:
Iterate as in scrum sprints. Have version control for your prompts – V1, V2, V3 etc, along with notes on what changed.
Specify your Audience:
Make the target audience clear to the AI. For example “ Explain a reasoning LLM to an 8 year old” yields dramatically different explanations and vocabulary as compared to “Explain the intricacies of a reasoning LLM model to a data science PhD student”.
Think in Chain of Tasks:
Break down complex tasks into smaller sequenced steps to help LLM navigate complex queries. The output of one task becomes the input of another. For example, search the internet for the most reliable sources on EHR software in home health> Search those sources for the most valued features in Home health EHRs> Prioritize the feature list according to RICE framework etc.
Taking a Step Back
AI or LLMs by themselves are not a threat that needs to be viewed with suspicion. But it is a capability that will profoundly enhance human productivity in almost every sector in future. The ones who don’t learn to use will inevitably be phased out by the market. Tough, but true. No one can guarantee you security or safety in future. But you will be far better positioned to play in the AI market of the future, if you have acquired the AI skills and have positioned yourself to take on challenges that will inevitably come up with this transition. As a tech-professional, you are in fact, in the best position to take it on. If you haven’t already, start here. All the best!
n