The five parts of a dependable prompt
Start with the task, add context, define constraints, specify the output format, and include a quality check when the task is consequential.
- Task: what should happen?
- Context: what does the model need to know?
- Constraints: what must or must not happen?
- Format: how should the answer be structured?
- Evaluation: how will you review it?
Before-and-after example
Weak: Write a product description.
Stronger: Write a 120-word product description for first-time buyers of a compact travel tripod. Emphasize stability, folded size, and setup time. Use plain English, a short opening benefit, and three bullet points. Do not invent specifications.
Iterate with evidence
When the result misses the mark, change one variable at a time. Keep the original goal visible, ask the model to identify assumptions, and compare the revision against a simple checklist.
A practical engineering loop
Begin by framing the task in one sentence. Add only the context that changes the answer, then define constraints and output shape. Run the prompt on a small test set. Review the result against the goal, record the failure, change one instruction, and test again. The loop is valuable because it gives you evidence for each revision.
- Frame: define the outcome and audience.
- Specify: add inputs, limits, format, and evaluation.
- Test: use normal, incomplete, and edge inputs.
- Review: separate instruction failure from missing data.
- Iterate: make one controlled change and compare.
Evaluate five dimensions
Score correctness, completeness, instruction-following, usefulness, and consistency separately. A fluent answer may still be wrong or incomplete. For factual tasks, add a verification step. For creative tasks, judge whether the output preserves the brief rather than whether it uses fashionable language.
- Correctness: are claims supported by the allowed inputs?
- Completeness: did the answer cover every required part?
- Control: did it follow format and limits?
- Usefulness: can the reader act on it?
- Consistency: does it behave similarly across inputs?
Fix the right failure
If the model invents a detail, improve the evidence boundary. If it ignores a required field, make the output schema visible. If the result is generic, provide audience and source context. If it is inconsistent, reduce optional instructions and add a small test set. Avoid adding a long list of negatives when one clear rule explains the failure.
Document the prompt version
Keep the prompt, model, date, test inputs, rubric, and observed failures together. This makes a good result reproducible and makes a regression visible when the model or source material changes.
Six prompts to try and adapt
Structured extraction
Extract product claims from the supplied page into a table with claim, supporting sentence, confidence, and missing evidence. If a field is absent, write unknown rather than infer it.
Adapt it: Change the fields and define what evidence is acceptable for your domain.
Controlled rewrite
Rewrite this paragraph for a beginner audience without changing facts. Preserve every number and named entity, then list any sentence whose meaning could be ambiguous.
Adapt it: Set the reading level and the facts that must remain exact.
Evaluation rubric
Grade the response below from 1 to 5 for correctness, completeness, format, usefulness, and consistency with the brief. Give one evidence-based reason per score and one revision priority.
Adapt it: Replace the rubric dimensions for creative, technical, or customer-facing work.
Missing-context check
Before answering, list the minimum missing information that could change your recommendation. Ask no more than four questions, explain why each matters, and wait for my answers.
Adapt it: Use this when the cost of a wrong assumption is higher than the cost of a question.
Test-set comparison
Run this prompt against the three inputs below. Return the output, the requirement it satisfies, and the first failure or ambiguity you notice. Do not rewrite the prompt yet.
Adapt it: Keep the test inputs representative and include one incomplete case.
Iteration plan
Compare version A and version B against the brief and rubric. Identify which single instruction changed the result, whether the change helped, and the next smallest experiment to run.
Adapt it: Add real versions, expected behavior, and the decision rule for keeping a change.
Frequently asked questions
What is prompt engineering in simple terms?
It is the process of writing and refining instructions so an AI system can understand the task and produce a result you can evaluate.
Does a longer prompt always work better?
No. Relevant detail helps; unnecessary detail can distract from the task and make the output less consistent.