If a model “does not get it,” the prompt usually left out something you assumed. Write that assumption down. A template is just those assumptions in a stable order.
Five fields
| Field | If you skip it |
|---|---|
| Task | The answer solves a different problem |
| Context | Tone and audience drift |
| Limits | The model invents numbers or policy |
| Output shape | You cannot reuse the result |
| Example | Style jumps around |
Output shape is the one teams skip. Ask for a table, a JSON object, or fixed headings if a script or a colleague will consume the text.
Copy this skeleton
Task: {what to produce}
Reader: {who it is for}
Limits: Use only the material below. If a fact is missing, write "unknown".
Material:
{paste}
Output: {exact headings or fields}
Example: {one short sample}Store templates by job (email, meeting notes, code review), not by product name. Products change. Add the date you last tested the prompt, and one failure you have already seen.
Tune one line at a time
Keep five real inputs. Change a single line, rerun all five, and write down whether the pass rate moved. If you change three lines at once, you will not know which one helped. A new teammate should be able to fill the braces and get an acceptable draft. If they still need a walkthrough, the example is too thin.
See a stricter version for meetings in A meeting-notes prompt.