The clause you did not write is the one that decides the answer. A model reads a request as a specification, fills every gap in it with the most ordinary thing that fits, and returns something fluent enough that the gap is invisible from outside.
Marek Wozniak runs the parts counter at Redbourne Tractors in Lincolnshire and spends half of every morning on the phone, reading fitment notes to farmers who are standing in a field. The notes are written for fitters: model codes, supersession chains, a serial cutoff. He asked a model to turn one of them into a paragraph he could read aloud.
The answer that dropped one line
The first ask was three words long, make this readable, and what came back was a clean paragraph naming the part, the tractor and the job the part does. It had quietly lost from serial 88410 onward. The farmer on the phone had an older machine, and the wrong part went in the post that afternoon.
Nothing failed in the way people expect failure to look. No part number was invented and no sentence contradicted another. A summary drops the qualifications first, because a qualification is not the subject: shortening keeps what a passage is about, and a serial cutoff is not what the passage is about, it is the thing that makes the rest of the passage safe.
Which sets the reading habit for every first answer. Read the first answer for what is missing, not for what is wrong. Wrongness announces itself on a second pass, and absence never does.
The request that keeps the line
Here is what Marek asks now, whole.
- Rewrite the note below for a farmer standing in a field who will hear it once, over the phone.
- The note is the only source. Do not add a part, a machine or a fitting step that is not in it.
- Every part number, every serial cutoff and every word limiting which machines this fits must appear in the rewrite unchanged.
- Return three sentences: what the part is, which machines it fits, what to check before ordering.
- If the note is unclear about which machines it fits, print that line unchanged and say it is unclear, instead of choosing.
The opening line names a reader rather than a tone, which is the substitution most templates make and lose by. Standing in a field, hearing it once, rules out subordinate clauses without anyone having to ask for plain English. It is the same move that being understood, in every form the job takes is built on, and worth the detour because a brief that works on a person works here for the same reason.
The second line is a fence. Fence the material and forbid anything that is not in it, and the commonest failure on a rewriting task stops being available at all: a model holds a great deal of general knowledge about tractor parts, and every sentence of it is a sentence Marek cannot check against the note in front of him.
The third line is the one that was missing in the morning. Name what must survive the rewrite, not only what to do with it. A rewrite instruction says what to change, and with nothing saying what may not change, everything in the note is in scope, so the qualifications go first because they are the least of the subject.
The fourth settles the form before the task, and the fifth is the escape hatch. Give the model somewhere to put what it is unsure about, or uncertainty has nowhere to go except into a confident sentence, because every exit from the request except answering it has been closed.
What a gap gets filled with
Every unstated thing in a request has a default, and the default is whatever was commonest in the text the model was fitted to. Unstated length arrives medium. Unstated audience arrives as a general reader. Unstated certainty arrives certain. None of those is a fault in the machine and none of them is announced on the way past.
That is also why a second draft is worth more than a folder of first ones, and why it only pays if it is done in single edits. Change one clause at a time, or you learn nothing from the second answer. Move three things at once and the answer does change, and which of the three moved it is now permanently unavailable.
Prompting
Five degrees of answering a question that was not quite asked
Every ring reads well. Nothing in the wording of an outer one tells you which ring you are holding.
- About the subjectTrue of the field, no use for the case
- The question as anyone asks itThe average of every version of it
- Your question, in general termsRight shape, none of your material
- Your question, on your materialThe first ring you can check a line of
- In the form you will use it inA rewritten note, not a description of one
The rings are degrees of missing, not degrees of quality. The outermost answer is usually the best written one, because it is the answer the model has produced most often.
The order of learning turns back on itself
Start with the four things a request needs before it works, then go straight to deciding whether to trust what came back, and then back to the first one with one clause changed. That is the whole loop, and it is the part that gets skipped, because rewriting the output by hand is faster on any single afternoon than rewriting the request.
Only after the loop turns does the rest pay. What the model can see when it answers is a question you cannot even ask before you have watched an answer go wrong for a reason. Several small asks that beat one large one comes after that, and taken first it is the wrong lesson: a person who splits a request before knowing which clause was failing ends up with four requests that each fail in the same place. The five failures everyone meets, and the fix for each reads as a glossary at the start and as an explanation once each of them has happened to you.
The stall is a folder of prompts
People stop making progress here at the point where they begin collecting. A prompt that worked once gets saved, pasted in next month, and edited only by adding, so the folder grows and the skill does not. Nothing in the folder records which clause did the work, because no clause was ever removed to find out.
The tell is a request that is long and does one job. Marek's is five lines, and every one of them was added after an answer went wrong in a way that line now prevents. A prompt is a draft of a specification, and a specification nobody has ever cut a line from has not been tested.
What to open in 2026
Google AI Studio is the free place to see a request whole. It shows the system instruction and the sampling controls that a chat window keeps out of sight, so the thing you are editing is the actual request rather than the visible half of it, and it replaced guessing what the chat window had put in front of what you typed. That matters for exactly one reason: two answers that differ can differ because you changed a clause or because the sampling did, and there is no way to tell those apart from the answers.
The saved instruction slot, which is Projects in Claude, custom instructions in ChatGPT and Gems in Gemini as of 2026, replaced pasting the same three lines of standing context at the top of every conversation. Which of the three carries it on a free tier has moved more than once, so check the tier before building a habit on it. The mistake is putting the task in there. What belongs in a standing instruction is what is true of every task you bring, and Marek's is one sentence about who is on the phone.
For keeping prompts where they can be compared, the llm command line tool logs every request and every response to a local database you can query afterwards, free and open source as of 2026. It replaced the chat history, which cannot be searched by what you changed.
Which leaves the question this page cannot answer. Marek can hold the rewrite against the note and see the missing line, and most requests worth making arrive with no note to hold them against. The checking then has to come from somewhere else, and the nearest source is enough of the mechanism to predict the behaviour, because it says in advance which parts of an answer were produced rather than recalled. The question gets sharper rather than easier once a model given tools and allowed to act is doing the asking, since the clause nobody wrote stops costing a paragraph and starts costing an afternoon.