Not every rule belongs in the prompt: steering AI instructions and model memory
The models get smarter, and yet around them grows an ever thicker layer of instructions. Exactly there sits a paradox. The better the AI works on its own, the more external instructions we put in its way, and because those are rarely cleanly written, they confuse the model as often as they help. An example of the sheer quantity: from Claude Fable 5, before the lockdown, the complete system prompt was pulled, around 17,000 words the model reads before you have typed a single word.
This system prompt is only the first layer. After it the model reads your own instructions, then the memory it keeps of you itself, and if you work in a project, additionally the project instruction. The system prompt is written by experienced engineers. Everything else comes from you, or the model guesses it together, with errors. Out come skewed answers, unnecessary follow-up questions and dialogues you have to restart.
The decisive question in personalization is therefore not “what do I write” but “where does it belong.” And the second: is what the model remembered about me still true? How I solve this in ChatGPT, Claude and the agents is here.
Six zones, and where what belongs
A principle up front: the broader an instruction acts, the more carefully you write it. An error in the one-off prompt spoils one answer. An error in the settings spoils answers over months, and you can hardly connect the odd behavior to a rule you wrote half a year ago and long forgot. Perhaps you did not even write it yourself; the model erred while summarizing.
There are up to six such zones. I show them on ChatGPT and Claude; other interfaces arrange it differently, the principle stays.
The first zone is the product settings. In Claude under Settings, then General and Capabilities; in ChatGPT above all the “Personalization.” Everything for which the provider has already built a switch: the model’s character, verbosity, thinking mode. The rule for this is simple. If the switch exists, use it instead of describing the same thing in text form. The ready-made switch is phrased by the developer and tested on millions of users, it takes hold more cleanly than your own phrasing.
The second zone is the user instructions. In ChatGPT the “About you” field plus the actual instructions, in Claude the field for instructions. Important: these instructions apply to all chats at once, including the ones already running. Write “answer concisely,” and the model compresses even where you need a detailed analysis.
The third zone is the project: a folder of chats to which you give your own instructions and add files with context and examples. In ChatGPT you additionally set whether the model accesses your entire memory or only the project’s. That is handy. If someone shares the account with you, create them their own project, turn off the shared memory and write into the instruction who works there.
The fourth zone is the memory: facts about you the service collects itself. A great function, at the same time the trickiest, because the model fills this field itself, often with errors.
The fifth zone exists fully so far only at Anthropic: skills. Your own guides for a recurring type of work that switch on only when they are needed. A fixed research process, an editing routine, a fact-check, described once and afterwards called by keyword instead of typed anew each time.
The sixth zone is the current prompt. Everything one-off lives here and only here: the one-off role, the one-off format, today’s task.
One piece of advice holds the table together: if you regularly write the same thing in the prompt, it belongs in a more permanent zone. That is how I handle my own knowledge. Per work context one folder with its own instructions and example files; alongside it a second brain in which only the permanent lies: methods, frameworks, tested prompts, sources, templates, findable via an index file. What belongs to a single client job stays in the respective project. Into the second brain migrates only what I reuse. The same hygiene as with the six zones, only at file level.
When zones contradict each other
The closer an instruction stands to the current task, the more heavily it weighs. In ChatGPT, project instructions officially override the global ones. The request in dialogue beats both, including the chosen “personality”: ask for dry answers, and the friendly preset switches off for this dialogue.
A tricky case remains, and OpenAI names it in the help itself. A saved memory can override a personality setting. Weeks ago the model noted “prefers factual, professional answers,” and your cheerful preset shows no effect anymore. The annoying part: it happens silently. You turn the switch, nothing happens, and that a line in the memory is to blame you can hardly guess.
So keep the conflicts small from the outset. They almost always arise where someone writes huge walls of text and tries to “program” every step of the model: how to answer, when to stop, when to ask back, which properties to have, what on no account to do. In early AI generations this was necessary, they floundered without clear instructions. Today’s models are trained on so many examples that they get by with very plain prompts. Big walls of text also cost context; where the tokens go I have shown elsewhere. The current guides from OpenAI and Anthropic say it directly: describe the desired result, not the path to it. The larger frame on this, soft presets and the gaps only you can fill, you find in the piece on Fable 5’s system prompt.
The user prompt is a contract, not a constitution
If walls of text do not work, what do good user instructions look like? Treat them as a short contract with the model, not as a basic law for every situation. Three rules help.
First: careful with “always” and “never.” The OpenAI prompting guides advise reserving absolute phrasings for the real prohibitions, that is what may happen under no circumstances. For the rest, conditions work better. “Always ask follow-up questions” interrogates you even where everything is clear. “Only ask back when the answer would be noticeably worse without the info” brings the follow-up exactly when it is needed. Such nudges belong to the techniques with which you get more out of every answer.
Second: one example beats ten prohibitions. If the model misses a format, do not extend the prohibition list in the instructions. Put a sample of a successful result into the working prompt. Rebuilding a structure is easier for the model than assembling it from rules.
Third: instructions are not a material store. Rules of behavior belong in the instruction; material the model should draw on, that is style guides, glossaries, work samples, belongs in the project files or the concrete dialogue. Whoever mixes both gets the worst of two worlds: the instruction bloats, and the reference work still sits in the wrong place.
Assembled, the scaffold of good instructions looks roughly like this:
About me: who I am and what I work on, usually two or three main tasks.
Style: language, form of address, no bureaucratese, tables only when they really help the comparison.
Collaboration: if information is missing, make a justified assumption and name it; on criticism, do not soften anything out of politeness.
What not: do not turn any answer into a lecture, do not fake a fact-check without having opened the source.
Your fill looks different, but the order of magnitude counts: 600 to 800 characters, not 6,000.
For contrast, the counter-example as it stands in many “best ChatGPT setting” lists:
“You are a brilliant world-class expert. Always answer as extensively as possible. Always joke. Always ask back. Never use lists. Never refuse. Be my best friend, lawyer, doctor and editor at once.”
Here almost nothing is right. Five absolutes on seven short sentences. The demands contradict each other: “as extensively as possible” goes badly with constant follow-ups. And the whole getup applies to every chat, to the dinner recipe as to the serious case. The newest generation (Claude Opus 4.8, Sonnet 5, GPT-5.5) now often ignores badly written prompts or points out the contradictions. A safeguard, yes. You should not rely on it: the model burns time with corrections, and the answer gets weaker in the end.
And if you give the model a personality?
Almost everyone has experimented with it. Understandable, preferences differ, some like it wound-up and witty, others the calm cynic. Only there is a trap in it. Humans are flexible: the joker knows when to keep quiet. Models are tuned to instruction fidelity, and so the AI comedian jokes even when it does not fit at all. This point I already opened in the system-prompt piece; here the practical side.
Look first at the ready presets. ChatGPT now has a whole set of tested characters in the settings: professional, friendly, efficient, cynical and more. Each takes hold at the level of the system prompt, with the same tool the developer uses to tune the model. You change the tone, not the way of thinking. If you want a cynic, take the ready one: tested on millions of dialogues, it tips into rudeness less often than one you describe in your own words. Claude has no such presets but, from the leaked system prompt, is recognizably set to a warm, friendly tone, for most cases the best setting.
If you do want to experiment, do not describe a personality in detail but the preferred tone as a range. “Write vividly, light irony is welcome in everyday and creative tasks; on health, money and conflicts stay neutral.” That leaves the model the room to choose the register by the situation. A rule of thumb on this: temperament into the settings, the professional role into the task. The strongest personalizer is at the same time the simplest, namely a name. Give the model one. The effect on the experience is big, the one on the answer quality practically zero.
Memory: the notebook that ages
Now to the most underrated zone. About memory two misconceptions circulate. Some believe the model remembers every conversation, others that it remembers nothing. The truth lies between. The memory consists of two mechanisms: the search over your chats and short notes into which the model enters what it considers important about you.
The chat search is an ordinary text searcher. The problem: the model does not always hit the moment in which it should search. Nudge it. “Read all chats in which we worked on topic X, gather the context and then search the web for what’s newly been added” pulls the existing context out more reliably than the hope that the model comes to it on its own.
The second mechanism is the notes, in ChatGPT “saved memories”. With the chat history, ChatGPT does not remember every detail. So do not hope it pulls the decisive thing out of a three-month-old dialogue on its own; the important you have saved explicitly as a note. The most inconspicuous fact about ChatGPT’s memory: deleting a chat does not delete the note from it. The notebook lies separate from the conversation. You delete the botched dialogue in which you wrote something private in anger, and the note about it stays and keeps seeping into the answers. Cleaning up you have to do separately: in the settings under Personalization, or in the chat with “forget that I …”. And turning off the memory does not delete what has accumulated, it only stops using it.
Claude follows a different philosophy. Instead of a scatter of individual notes, a summary: Claude reassembles a concise portrait from all conversations about once a day. Project chats do not flow into the general portrait, each project has its own memory (you remember the trick with the project for shared use?). And the focus is narrower: Claude deliberately remembers the professional, that is role, projects, style, technical preferences, and rather drops the everyday. You can look at and edit this under Settings, Capabilities, Memory, or you say directly in the chat what to remember and what to correct; the change takes hold from the next conversation.
That is the reason I do not leave the permanent to the model’s guessing in the first place. What holds long stands in my second brain, with a date and controlled by me, separate from what the service notes itself.
Memory review: ten minutes a month
The memory has an unpleasant property: it does not age with you. You change jobs, the model knows the old position. You finish a project, it keeps appearing in the answers. Every outdated note mixes in quietly, and the longer you use the service, the thicker this sediment gets.
I have therefore adopted a monthly review, it costs ten minutes. Open the memory (in ChatGPT the memory management, in Claude “Memory”), skim it by eye, then give the model two prompts:
“Read what you’ve saved about me and assess which information you’re missing. Ask me questions, three at a time, with the most important first; from my answers you update the memory.”
“Read what you’ve saved about me and assess which information might be outdated or incomplete. Ask me questions, three at a time, with the most important first; from my answers you update the memory.”
If you work with two or more models, one more trick. Open a chat and write: “Read what you know about me and create my meta-card.” The meta-card you tip into the other model’s chat and have it align its memory to it.
And the agents?
How memory is organized in AI agents I have shown on the build of an AI agent. The storage looks a bit different, the principle stays: keep global instructions, project instructions and the model’s memories of you cleanly separate. Whoever keeps these three levels apart has already done the biggest part of the work.
The right place beats the longer prompt
At the start stood the paradox: the models get smarter, and the layer of instructions around them gets thicker anyway. The way out is not: longer prompt. It is: the right place for every rule, plus a memory that fits the you of today. Open the memory management of your most-used model now and read the first ten entries. Bet at least one is no longer true?
Frequently asked questions
Where does an instruction that should always apply belong? Into the user instructions if it concerns you as a person (language, form of address, role), or into the product settings if there is a switch for it. Everything that belongs to only one job stays in the project. Rule of thumb: the broader the effect, the more concise and careful the phrasing.
Does deleting a chat also delete what the model remembered? In ChatGPT, no. The saved memories lie separate from the chat history and outlast the deletion of the conversation. Clean them up separately, in the settings under Personalization or via “forget that I …”.
How often should I check the memory? Once a month usually suffices, around ten minutes. Old job titles, finished projects and outdated preferences otherwise accumulate and seep unnoticed into the answers.
Should I prescribe the model a detailed personality? Better not. Take a tested preset or give a tone range (“vivid in everyday, neutral on money and health”). A rigid detail-personality jokes even when it is inappropriate.
Does this apply to ChatGPT and Claude equally? The principle of the zones, yes. The mechanics differ: ChatGPT keeps individual notes and separates saved memories from the chat history; Claude builds a summary once a day and keeps the project memory separate.