You write to an AI assistant: cut this offer down to one page. Back comes a text half the length, with the paragraph about pricing taken out – the one paragraph that had to stay. Nothing broke here: the prompt never said a word about what to leave alone.
Today we look at what a request like that leaves to guesswork, and when there is nothing to add because one sentence already does the job.
Claude Code fills in whatever the prompt leaves out
Claude Code’s documentation puts it plainly: Claude can infer your intent, but it can’t read your mind. Anything that isn’t in your prompt and isn’t in the files sitting in your folder, the AI assistant supplies for itself. It fills in the most likely version, and yours is often a different one.
Anthropic’s prompting guide offers a comparison worth keeping: treat Claude as a new employee who is brilliant at the job and doesn’t yet know your norms or the way you work. It has no idea who your clients are, and no idea that the pricing paragraph in this offer is untouchable.
One test settles most cases: read your prompt as if it were going to somebody who isn’t in the middle of this with you. If that person would ask “which offer?” or “how short?”, the AI is going to guess in the same spot.
When one sentence is enough
Before you add anything, check whether there is anything to add. A prompt you can put in one sentence – fix the typo in the second paragraph, put today’s date in the file name, delete the empty section – has nothing to make more precise. The scope is obvious and the result is visible at a glance.
There’s a second case where a loose prompt is exactly the thing you want: reconnaissance. “Look through this folder and tell me what’s worth tidying up here.” You don’t know yet what you want to ask for, so narrowing it down would take away the very thing you came for – an answer you weren’t expecting.
On a bigger job, point at the material and at an example
What a bigger job is missing most often isn’t precision, it’s material. “Write an offer for a new client” leaves the whole lot to guesswork: your prices, the scope of what you sell, the way you usually describe it. Claude Code has access to the files in the folder you start it in, so rather than copying prices into the prompt, point at the file they already live in.
An example covers the other half. Instead of describing in words what you’re after, point at a document that already looks that way – one of your earlier offers, or a note written in your own voice. The same prompt with material and an example looks like this:
Write an offer for a new client. Take the prices and the scope of services from my price list, and copy the layout and the tone from an earlier offer – both files are in this folder. Before you start writing, tell me what you're missing.File names are nothing you have to remember. Claude Code goes through the folder and finds them. If several similar price lists are lying in there, name the right one in the prompt. That request for questions belongs in every bigger prompt: a few questions before the work cost less than fixing a finished text.
Say what you need it for
The reason is the easiest part to leave out, because to you it’s obvious. “Cut this offer down to one page” describes an action. “Cut this offer down to one page, the client reads it on a phone right before the meeting” describes an action and a reason. That reason settles all the small decisions the request says nothing about. Lose the price list, or lose the paragraph about the company? With the reason in view there’s only one answer.
Anthropic’s prompting guide says the same thing: give the context or the reason, because AI can generalise from that explanation. The prompts we hand clients at our workshops carry a heading at the very top – GOAL – and one sentence under it.
Add one thing that can be counted
Claude Code stops working when a task looks finished. On a longer job that matters: if nothing in the prompt tells you the result is good, “looks finished” is the only signal in play, and the whole job of checking lands on you.
The criterion doesn’t have to be technical. “Three questions in the FAQ”, “one page at most”, “company names stay as they are”, “one thought per paragraph” – every one of those can be counted or seen on the spot. Put it in the prompt and ask Claude Code to check it right away: “at the end, check there are exactly three questions and tell me how many you got”.
The whole cycle of working with Claude Code, from the plan to checking the result, is in our lesson on the first working session.
Ask why it came out differently
The result isn’t the one you wanted, and the first reflex is to write: no, wrong, fix it. The fix usually lands, except that on the next job of the same kind it all comes back: one bad result is gone, and the reason behind it is still sitting in your prompt.
Instead of criticising the bad result, ask first where it came from:
Before you fix it: why did you do it this way? How did you read my prompt?The answer shows what Claude Code took from your request, and most of the time it turns out it read exactly what was there. “Cut the offer” with nothing said about the price list means “cut the offer”, and the price list is the longest thing in it, so out it goes first, logically enough.
We start with “why?” ourselves on every bigger job whose result surprises us. It usually ends with one sentence added to the prompt, and that mistake doesn’t come back.
When you ask again for the same thing, write the prompt from scratch
Maybe you’ve had a spreadsheet that grew for so long, and got so tangled along the way, that at some point the best move was to build a new one from scratch.
AI is much the same – when a conversation runs on too long, it can start going round in circles, returning to the same wrong answers instead of reaching different, better ones. You’ll recognise the moment by this: you keep asking your AI assistant for the same correction over and over, and every result is worse than the one before.
Claude Code’s documentation suggests clearing the conversation at that point with /clear and starting again, carrying everything the failed attempts taught you. The reason is the same as with the spreadsheet: every rejected version is still in that conversation, and Claude Code reads all of it, not just your last request.
Is the job big enough that you don’t know what belongs in the prompt? Turn it around and let yourself be interviewed:
I want to tidy up my folder of document templates. Before you do anything, ask me about the details – one question at a time. Ask about the things that are easy to overlook, too.In short
On a small job you can judge at a glance, one sentence is enough. Adding anything to it is time you won’t get back. On a bigger one: point Claude Code at the material it should work from, say what you need it for, and throw in one thing that can be counted at the end. Result isn’t right? Ask why it came out that way before you criticise it. And when you’re asking again for the same thing with nothing to show for it, open a fresh conversation and write a better prompt instead of correcting it one more time.
Prompting AI is a bit like casting spells: what counts is the words you actually use, not what you meant. A spell doesn’t know your intentions – it does exactly what you tell it. Leave one thing unsaid and the prince stays a frog for good. Happily, with AI every spell can be cast again as many times as you need, and nobody takes offence.