💭 in your head · nothing to type
Picture a genie — and you're in the middle of negotiating the terms of a wish. The genie grants exactly what you say, nothing more, nothing less. It's a trickster: any loophole in the wording and it will find it.
Run it in your head: surface the task that repeats every week for you — a report, meeting prep, analysis, content, replies. Hold it. Now picture bargaining with the genie over the exact wording of this wish. What do you need to nail down — so the genie can't wriggle out?
💭 in your head · nothing to type
Now picture a different scene: you're a theatre director blocking a scene before the first rehearsal.
Hold your recurring task in your head as that scene: who's in it, what their lines are, how the act ends. A director doesn't trust that actors will figure it out — they set the positions in advance.
Notice: which "positions" in this task have you never set for AI? What stays in the wings if you don't put it on the page?
If you don't know where to start — ask Claude to write the prompt for you:
You are an expert at writing prompts for language models.
Write a detailed prompt for an assistant that will [task description].
Each of its responses must:
- [criterion 1]
- [criterion 2]
- [criterion 3]
Include all parameters important for quality.
Sometimes it works better from the other end. First solve the task together with the model in an ordinary chat: clarify, correct, until the result is what you need. Then ask the model to pack that path into a prompt:
We have just brought this task to a result I like.
Write a prompt that gets the same result on the first try
with similar data. Include everything I clarified and corrected along the way.
Test the prompt in a new chat on a different example. If the result is worse, add what was missing and test again. A prompt like this already knows your corrections instead of guessing them.
For repeating tasks (Claude Project, Gemini Gem, Claude Code Skill):