The five levers, and why they work

A vague request gets a vague response. That's not inevitable: five levers are enough, and you never need all five at once.

1. The role

"You are a demanding proofreader." It looks like a game, but it's actually the most effective lever. By naming a role, you steer the model toward a specific region of what it has read: the vocabulary, reflexes and priorities of a proofreader rather than those of a generalist.

2. The context

"It's for a client I don't know, in construction." Without it, it answers the average question asked by an average person. Context is what turns a correct answer into a useful one.

3. The format

"In five bullet points, no introduction." The built-in flaw of these tools is length. If you say nothing, you'll get three paragraphs of preamble before the actual answer.

4. The constraints

"Maximum 150 words. No jargon. Don't suggest anything over 50 euros." A negative constraint is often more effective than a positive one: saying what you don't want frames things better than saying what you do.

5. The example

"Here's a message I liked, write one in the same vein." The most powerful lever, and the least used. An example conveys a tone, a rhythm and a level of detail that three paragraphs of instructions can't describe.

The right balance

Don't pile up all five out of habit. An overloaded request becomes contradictory, and the model then arbitrates on its own, usually badly. Two well-chosen levers beat five stacked together.