How to write a better prompt
the five things every prompt needs
A prompt usually goes wrong because something is missing, not because a sentence was clumsy.
Download for Mac
- Every prompt needs five things: a goal, context, steps, limits, and how you’ll know.
- Context is what the agent can’t see: your screen, the error, what you tried.
- Typing it all is slow. Say it, with a screenshot, in one go.
You type one line. The agent guesses the rest, and you spend four messages correcting it.
Five things every prompt needs
- 1Your goalWhat should be true when it is done.
- 2The context it can’t seeThe error, your screen, what you tried.
- 3The stepsWhat to do, in order if order matters.
- 4What not to touchThe part everyone forgets.
- 5How you’ll know it workedSo it can check its own work.
Fix the onboarding redirect for Google sign-ups.1 New users hit the dashboard before their workspace exists, so it 404s. I already tried a retry, it didn’t help.2 Fix the race in the callback handler, then add a test for a fresh Google sign-up.3 Don’t touch the email sign-up path.4 It’s done when a new Google account lands on its own dashboard.5
No headings needed. A prompt goes wrong when one of the five is missing, not when it is in the wrong place.
Context is what the agent can’t see
- ✕No: Your screen: the broken page, the odd layout.
- ✕No: The error in your browser or your logs.
- ✕No: What you already tried, so it doesn’t try it again.
- ✕No: Why the odd-looking code is there.
So show it. A screenshot of the broken page says more than a paragraph about it.
Four rules that hold up
- ✓Yes: Say what not to do. Cheap to write, expensive to leave out.
- ✓Yes: Name things exactly.
src/main/pipeline.ts, not “the pipeline”. - ✓Yes: One example beats three adjectives.
- ✕No: Skip the padding. Politeness and role-play preambles buy nothing.
That is a lot to type
- ✕No: Five parts take longer to type than the question.
- ✕No: Then you save the screenshot, find it and drag it in.
- ✕No: The “don’t touch that” is the part you forget.
Wouldn’t it be so much easier to just say it?
Claude Code /voice
Wispr FlowMaybe add a CSV option.
Take the screenshot, then talk
With Yappr, a screenshot you take lands on your notch. Hold your key and say what is wrong, and it goes in with your words: in Claude Code, as a file it can open. Yappr keeps your own sentences, in the order you said them. It does not add headings.
Speaking into Claude Code? Compare /voice with a dictation app.
The same rules belong in the standing file: what to put in CLAUDE.md.
Written 6 September 2026. Every fact about another product is from its own pages, checked that day.