When output is poor, most people rephrase and hope. The skill worth building is being able to say precisely which element was missing — and add it deliberately.
Why vague prompts produce vague output
An AI model has no access to your situation. It does not know your role, who will read the result, how much detail you need, what format you require, or what "good" looks like to you. When you leave those unstated, the model does the only thing it can: it produces the statistically average response to your question.
Average is exactly what you get. Generic, plausible, mildly useful, and not quite right. Then you conclude AI is overrated.
Consider the difference:
Vague
Write about project management.
Produces a generic overview useful to nobody in particular.
Specific
You are a delivery manager. Write a 200-word update for non-technical executives explaining a two-week slip, its cause, and the recovery plan. Direct tone, no jargon, end with what we need from them.
Produces something you could almost send.
Same model. Same session. The entire difference is supplied information.
The five elements
Context — the situation and background
What is happening, what led to it, and what the model needs to know to respond sensibly. This is the most frequently omitted element and the most costly one.
Weak: "Help me write an email."
Strong: "A client missed our third deadline. We have worked together two years. I want to raise it firmly without damaging the relationship."
Level — the required difficulty or experience level
Tells the model how much to assume. Without it, explanations land either patronising or impenetrable.
Examples: "Explain as if I have never seen this before." · "I have eight years in this field, skip the fundamentals." · "Pitch it for a smart colleague from a different department."
Expected outcome — what you actually want
Name the artefact. "Help me with the report" is a topic. "Produce a one-page summary with three findings and a recommendation" is an outcome.
A useful test: could someone else look at the result and tell whether it satisfied your request? If not, you named a topic rather than an outcome.
Audience — who the result is for
Changes vocabulary, length, tone, and what can be assumed. The same content for a CEO, a junior engineer, and a customer are three genuinely different documents.
Examples: "For our CFO, who cares about cost and risk." · "For new joiners in week one." · "For a frustrated customer awaiting a fix."
Requirements — format, tone, limits, quality standards
The constraints that make output usable without reformatting. Length, structure, tone, things to include, things to avoid.
Examples: "Under 300 words." · "Use a table." · "No marketing language." · "Include one worked example." · "Flag anything you are uncertain about."
For simple requests, Context and Expected outcome are often enough. The framework is a diagnostic checklist, not a mandatory form. Its real value appears when output disappoints — run through the five and you will almost always find the missing one.
The reusable template
That final instruction matters more than it looks. Asking the model to review its own output against named criteria consistently improves the result, because it forces a second pass rather than a single generation.
Diagnosing bad output
This is where the framework earns its place. When a result disappoints, do not rewrite blindly — identify the symptom.
| Symptom | Missing element | Fix |
|---|---|---|
| Generic, could apply to anyone | Context | Add your situation, constraints, and history |
| Too basic, or too advanced | Level | State your experience explicitly |
| Interesting but not what you needed | Expected outcome | Name the exact artefact and its structure |
| Wrong tone or wrong vocabulary | Audience | Say who reads it and what they care about |
| Right content, unusable shape | Requirements | Specify length, format, and what to avoid |
| Confidently wrong facts | Requirements | Add "flag anything uncertain" — then verify independently |
Before you press send
- Did I explain the context?
- Did I clearly state the outcome I want?
- Did I identify the audience?
- Did I mention the required format?
- Did I define important limitations?
- Did I ask for examples?
- Did I ask the AI to review its own answer?
A worked upgrade
Watch the same request improve across three passes.
Returns generic advice: list achievements, be positive, ask about growth. Nothing you did not already know.
Substantially better — now specific to the situation.
Now returns something genuinely usable — specific, defensible, tailored to the person in the room.
The third prompt took perhaps ninety seconds longer to write. It saved an hour of editing and produced a materially better outcome.
Common mistakes
Practise this week
✍️ Prompt practice
A note, a doc, anywhere you will find it again.
What comes next
You now have the structure. The next component gives you working prompts built on it — ready to adapt for learning, career planning, productivity, email, meetings, research, and problem-solving.