A Practical Prompt Framework That Works
Prompting is not a secret language. A good prompt is a clear work request. It tells the system what outcome you need, what information matters, what boundaries apply, and how the result will be reviewed.
The six-part framework
- Goal: What should be produced or decided?
- Context: What background information changes the answer?
- Constraints: What must the output include, avoid, or respect?
- Format: How should the answer be organized?
- Examples: What does a suitable result look like?
- Verification: What should be checked, cited, or marked uncertain?
1. State one clear goal
“Help with marketing” is too broad. “Create three headline options for a landing page promoting a bookkeeping service to small restaurants” is testable. If a request contains several unrelated goals, split it into stages.
2. Provide decision-relevant context
Context is not every fact you know. Include only information that should change the output: audience, location, product, current situation, skill level, deadline, or existing policy.
3. Define constraints
Constraints prevent polished but unusable output. Examples include a word limit, banned claims, required tone, legal disclaimer, supported software version, available budget, or source requirements.
4. Choose a useful format
Ask for a table when comparing options, numbered steps for a procedure, a checklist for review, or JSON only when another system will consume the result. Format should make the next action easier.
5. Give examples carefully
Examples help establish tone and structure. They can also cause imitation. State what should be copied—the level of detail or organization—and what should not be repeated.
6. Build verification into the prompt
Add instructions such as:
- “Do not invent sources or quotes.”
- “Label assumptions and time-sensitive claims.”
- “Show the calculation before the result.”
- “Ask one question if a missing detail would materially change the answer.”
Before and after
Before: “Write a customer reply.”
After: “Draft a calm reply to a customer who says a rented vehicle was delivered late. Acknowledge the delay, do not admit facts we have not verified, ask for the booking number, explain that the team will review the timeline, and keep the reply under 130 words. Output only the message.”
Use stages for difficult tasks
For a complex report, do not ask for the final document immediately. Ask the AI to:
- Summarize the objective and constraints.
- List missing information.
- Propose an outline.
- Draft one section.
- Run a quality checklist.
This makes errors visible earlier and reduces the cost of rewriting.
A reusable template
Goal: [outcome]
Context: [relevant background]
Constraints: [limits and rules]
Format: [structure]
Example: [optional sample]
Verification: [facts, sources, uncertainty, calculations]
Do not over-prompt
Long prompts are not automatically better. Remove instructions that do not affect the result. When output is weak, diagnose the problem: missing information, conflicting constraints, unsuitable model, or a task that needs a tool rather than more wording.
The final test
A strong prompt produces output that is easier to judge. If you cannot define what “good” looks like, improve the task definition before asking the AI to complete it.
