AI Tools Hub
25 prompts · PDF

25 prompts that hold up at work

The 25 we reach for most, pulled from our library of 110: reviewing a diff, debugging from a stack trace, cutting a draft, summarising a meeting into decisions, and checking what a dataset actually contains. Each one fixes its output format, and several make the model ask before it answers.

Nothing here is hidden

Every prompt in the pack is also on this site, free and without an email address, in the prompt library. The PDF exists because a file you can keep open on a second screen is more useful than a page you have to find again.

What is inside

  • Coding6
  • UI and UX design4
  • Writing and copy4
  • Research3
  • Working with AI3
  • Summarising and structuring3
  • Data analysis2

Three of the twenty-five

NO.01ChatGPT / Claude

Review a diff for the failure modes, not the style

Review the following diff. Ignore naming and formatting entirely — I only want failure modes.

For each issue, give me:
- The specific input or state that triggers it
- What goes wrong as a result
- The smallest change that fixes it

Rank by how likely the trigger is in production. If you find nothing above "theoretically possible", say so rather than filling the list.

<diff>
[paste the diff]
</diff>

When to use: Before asking a colleague to review, to clear the mechanical issues first

NO.02ChatGPT / Claude

Debug from a stack trace without guessing

Here is an error and the surrounding code. Before suggesting a fix:

1. State what the error actually means, mechanically
2. List every condition that could produce it, most likely first
3. For each, tell me the single cheapest thing I can check to rule it in or out

Do not propose a fix until I tell you which check came back positive.

<error>
[paste the error and stack trace]
</error>
<code>
[paste the relevant code]
</code>

When to use: When an error is unfamiliar and you are about to start changing things at random

NO.03ChatGPT / Claude

Generate tests from the behaviour, not the implementation

Write tests for the function below using [test framework].

Rules:
- Test observable behaviour only. Do not assert on internal calls or private state
- One test per behaviour, named so a failure tells me what broke
- Include the boundary cases: empty input, maximum size, null, and the error path
- Do not test the language or the standard library

List any behaviour you could not test and why.

<code>
[paste the function]
</code>

When to use: After the code works and before it goes near production