Transparent and reviewable

How AILesson prompt recipes are designed and tested

A transparent method for choosing useful tasks, writing model-portable prompts, testing real examples, and keeping limitations visible

1

Start with a real outcome

Every recipe begins with a task a person needs to finish and an output they can inspect. We avoid decorative personas and instructions that do not change the result.

2

Collect the minimum useful context

Variables capture information that only the user knows and that can materially affect the answer. Missing information remains an explicit placeholder instead of becoming a model guess.

3

Make uncertainty reviewable

Templates distinguish supplied facts, assumptions, evidence, recommendations, and open questions when the task requires it. Time-sensitive claims must be checked again at the time of use.

4

Test both languages with real examples

A recipe is not complete until its English and Simplified Chinese versions have full, non-sensitive input and example output. The public record names the tested tools and review date.

5

Check the result, not the prompt’s appearance

Each recipe includes observable checks for omissions, unsupported claims, changed commitments, and unusable output. A longer prompt is not automatically a better prompt.

6

Keep one source of truth

A recipe has one stable page even when it appears under several topics, audiences, tools, or workflows. Tool pages only include recipes whose published test record names that tool.