How to write a prompt that actually works

Founders imagine prompting is a skill with secret tricks. Mostly it isn’t. The gap between a mediocre prompt and a good one is almost always one of three things, in order of how often they matter:

1. Context. “Write our investor update” gets you a generic draft. “Write our investor update — here are our last three, and here’s this month’s actual numbers” gets you a real one. The single biggest failure mode is founders describing their situation in a sentence instead of pasting the actual thing.

2. Specificity about the output. “Make it good” isn’t an instruction. “Under 300 words, in my voice, three sections, flag anything you’re not sure is accurate” is. Say what “done” looks like.

3. Room to push back. Tell it explicitly to flag assumptions, or to say “I’m not sure” instead of guessing. Left alone, a model will often fill a gap with something plausible-sounding rather than admit it doesn’t know — asking for the flag is usually enough to get an honest one instead.

Everything else — exact wording, whether you say “please,” clever framing devices — matters far less than these three. If a prompt keeps giving you thin output, check which of the three is actually missing before you rewrite the wording.

This is the whole idea behind the memory files in the roadmap: they’re just a way of giving Claude real context by default, so you’re not rebuilding it in every prompt.