Twenty Minutes Once. Better Answers All Month.

You fix the same four things in every answer.

Too long. Too hedged. Written for a beginner when your readers are not beginners. Opens with a summary of what it is about to say, which you delete every single time. You fix them, you ship, and tomorrow you fix them again — because every session starts from nothing, and nothing is where you keep leaving it.

The tool has not failed you. You have just never written down what you want, and there is a version of that document you can build in twenty minutes.

Four steps.

1. Collect what you keep fixing

Here are the last five things I asked you, and what I changed in each answer before using it: [paste them].

Group my edits by what caused them. I want the four or five corrections I keep making, not a list of twenty.

Start from evidence, not from imagination.

If you try to write your preferences from memory you will produce a list of things you believe about your taste, which is a different list from the one your edits reveal. Everyone thinks they want "clear and concise". Your actual edits say something more specific and more useful — that you cut every sentence that begins with "In today's", or that you always move the conclusion to the top.

The grouping instruction is what makes this fast. Twenty individual edits are noise; four causes are a document.

2. Turn each one into a rule

Write each of those as a standing instruction, in the imperative, that would have prevented the edit.

Every rule must be checkable by looking at the output. If I cannot tell whether you followed it, rewrite it until I can.

Checkability is the entire quality bar here.

"Be concise" cannot be checked. "No paragraph over four sentences" can. "Write with authority" cannot be checked; "never use might, could, or perhaps about something you can verify" can. The uncheckable version is not a weaker rule — it is not a rule at all, because nothing can ever violate it, which is also why it changes nothing.

The imperative matters too. Descriptions of your preferences ("I tend to prefer shorter openings") get treated as background colour. Instructions get followed.

3. Add what it can't guess

Now ask me for the three facts about my work you could never infer from a request — who this is for, what I am not willing to say, and what "done" looks like.

Ask one question at a time. Do not accept an answer that could apply to anyone.

Rules cover how you want things written. These three cover what you are actually doing, and no amount of style instruction substitutes for them.

Who this is for is the one that changes the most per word. "Small business owners" and "solo founders who have done this before and are impatient" produce different work from the first sentence.

What you are not willing to say is the one nobody writes down. Every person has claims they will not make — guarantees, numbers they cannot stand behind, a tone they find embarrassing. Left unstated, those show up in a draft and you delete them, every time, forever.

What "done" looks like is what stops the endless polishing loop, in both of you.

The one-question-at-a-time instruction is not a nicety. A single message asking all three gets three shallow answers; asked one at a time, the second answer is usually better because the first question made you think.

4. Keep it honest

Assemble the rules and the three facts into one block I can paste at the start of any task.

Then tell me which rule you expect to break most often and why. That one is either badly written or it is the one that matters — and I want to know which.

The last question is the one that keeps this from rotting.

A rule it expects to break is a rule under tension: either it is vague enough to be unfollowable, or it cuts against a strong default and will need reinforcing. Both are worth knowing on day one rather than discovering in three weeks when you have quietly gone back to fixing everything by hand.

Then use it. Paste the block at the start of a task, or put it wherever your tool keeps persistent instructions — project instructions, custom instructions, a saved prompt, a file you keep open. The mechanism does not matter. Being first does.

What actually changes

Not the quality of any single answer — you could have got that by editing. What changes is that you stop paying the same twenty seconds of re-explanation forty times a month, and you stop accepting the drafts you were too tired to fix.

Add a rule whenever you catch yourself making a correction you have made before. The document is never finished, but it is useful the first day.

Stop re-explaining yourself every session.

---

The four prompts are in the carousel on [Instagram](https://instagram.com/sharpaihub).

Want the calm version of AI news like this, once a week? Subscribe to the Sharp AI Hub newsletter →