Rexz Bot logo Rexz BotBots and automation, explained plainly
AI Assistants

AI Assistant Custom Instructions That Are Worth Writing

Most people start every chat with an AI assistant by retyping the same three facts. I am in Sweden, so use metric and euros or kronor. Keep it short. Do not start with a compliment about my question.

A laptop showing a settings panel with a short block of typed preferences beside a notebook with handwritten test prompts

Most people start every chat with an AI assistant by retyping the same three facts. I am in Sweden, so use metric and euros or kronor. Keep it short. Do not start with a compliment about my question. After a month of that, the typing adds up, and the days you forget, the answer comes back in Fahrenheit with a paragraph of throat clearing on top.

Custom instructions are meant to fix this. ChatGPT, Claude, Gemini and Microsoft Copilot all have some version: a box where you write standing preferences that get attached to every conversation. Most people either leave it empty or fill it with a long wish list that the assistant half ignores. Both waste the feature. Here is what I put in mine, what I took out, and how to tell whether it is working.

What custom instructions actually do

The text you write is quietly added to the start of every new chat, before your first message. The assistant reads it the same way it reads anything else you type. There is no special enforcement and no hidden rule engine behind it. If an instruction is vague, it gets applied vaguely. If two instructions conflict, the assistant picks one, and you will not always like the choice.

This is different from memory. Memory features collect facts from your conversations over time and decide on their own what to keep, which we covered in our look at what assistant memory really remembers. Custom instructions are fixed text that you control completely. I trust them more for exactly that reason: I know what is in there, because I wrote it.

The boxes have limits. ChatGPT splits it into two fields of about 1,500 characters each, one about you and one about how to respond. Claude uses a single preferences field at the account level, plus separate instructions inside projects. The limits change, so check yours, but plan for roughly 200 to 300 words in total.

What belongs in the box

After about eighteen months of editing mine, I think good instructions fall into three groups. Anything outside these groups usually belongs in a single chat, not in a standing rule.

Facts that change answers

Your country, units, currency, language, and the tools you use daily. I write that I live in Gothenburg, use a Mac and Google Workspace at work, pay in kronor, and want dates written as 2026-03-14. That one line removes most of the corrections I used to type. If you work in a specific field, say so plainly, as in I run bookkeeping for three small cafes. The assistant pitches its answers differently once it knows that.

Format rules you would otherwise repeat

Keep these concrete. Write answers under 200 words unless I ask for more works. Be concise does very little, because every assistant already believes it is concise. Other rules I keep: no headings in short answers, tables only when I compare three or more things, and code blocks with the language named.

Things to stop doing

These earn their space more than anything else. Mine say: do not open by praising the question, do not close by asking whether I want more help, and if you are unsure of a fact, say so in one sentence instead of hedging every line. Negative rules are oddly effective, probably because they point at a specific habit rather than an abstract quality.

Longer is not better, and personas mostly backfire

A popular piece of advice says to give the assistant a detailed persona: you are a world class strategist with 30 years of experience, and so on. Prompt guides love this. In my testing it does very little for everyday use, and sometimes makes things worse.

I ran a small comparison last winter. I used the same 40 ordinary requests, things like drafting a polite reminder email, explaining a VAT rule, or rewriting a product description, against two versions of my instructions. One was my plain 180 word version. The other added a 250 word expert persona on top. I rated the answers blind. The plain version won or tied on 31 of the 40. The persona version was more confident, longer and more likely to give advice I had not asked for, which is the opposite of what I want from a daily tool.

Length has its own cost. The more rules you add, the more each one gets diluted, and the more likely two of them clash. Use simple language and give me full detail sounds reasonable until you ask a simple question and get an essay. When I trimmed mine from about 520 words to 180, the answers got better, not worse.

A good instruction is one you would notice missing. If you cannot remember why a line is there, delete it.

A version you can copy and adapt

Here is roughly what mine looks like now, rewritten slightly so it works as a template. Replace the details with your own and cut anything that does not apply.

About me: I live in Gothenburg, Sweden. I run the office side of a small design studio with six people. We use Google Workspace, Slack and a Mac. Currency is SEK, units are metric, dates are written year first.

How to respond: answer in under 200 words unless I ask for detail. Start with the answer, not with background. Use plain sentences, no headings for short replies. If a question has a practical step, give the step first. If you are not sure about a fact, say so once. Do not praise my question. Do not offer further help at the end.

That is about 100 words. It leaves room for one or two rules you discover you need later.

How to test whether your instructions work

Keep five test prompts in a note, chosen to hit each rule. Mine are a quick unit conversion, a short email draft, a comparison of two tools, a factual question I know the answer to, and a request for a spreadsheet formula. When I change the instructions, I open a fresh chat and run all five. It takes about six minutes.

Look for the rules that failed, not the ones that worked. If the assistant still opens with praise, the rule is probably buried too far down or worded too gently. Move it up and make it blunt.

Keep in mind that instructions apply to new chats. An old conversation will keep its old behaviour, which confuses people who edit the box and then continue a thread from last week.

For tasks you do repeatedly but not daily, such as writing a monthly client report, a standing rule is the wrong tool. Save the prompt instead. Our guide on keeping a personal prompt library shows how I organise those, so the custom instructions box stays short and general.

FAQ

Do custom instructions work across all my devices?

Yes, for the major assistants they are stored in your account and apply on web, desktop and mobile apps once you sign in.

Should I put private information in custom instructions?

Keep it to what improves answers. Your city and job are useful. Your address, ID numbers and client names are not needed and are better left out.

Why does the assistant sometimes ignore my instructions?

Usually because a rule is vague, conflicts with another rule, or conflicts with what you asked in the chat itself. A direct request in the message normally wins over a standing instruction.

How often should I update them?

I review mine every two or three months, or when I notice myself typing the same correction twice in a week.

HL
Hedda Lindgren

Hedda covers AI assistants and everyday automation, translating the hype into practical habits anyone can use.

More posts by Hedda

More in AI Assistants

01doorganico 084goad 3hot8 4465456 after5internprogram ahandfulofblessings ahmadpc ahomcybersecurity allodialresources annodex antiquesandperiodcollectibles aotc asianworldkashmir beedlectrics bloggertemplatefree bluefeatheragency bodyweightsolution campaign08blog carfixadvisor catstreetcollective cbstvsports cnfanspurchasing committeeof500yearsofdignityandresistance covertopsdjs cs-kaiyuangame cuestacomunicaciontotal cxoutlet dancebetgame deskgearhub destinyvillastay detectmagazine di33 diamondsteelbuildings donmo dyanadropship emarsysfinance emimattl escapechan eskitarih evautoplay everythingweddings-orlando farmzone feedviewerapp foxstudiosscratch gamevui123 gleasongrade greetingsday hacklabasu haskell iw-kaiyun jagoanproperti jilidskripsi joltbusinessfinance juegodepokergratis juegospucca kaeru-festival latestmehndidesigns lestanzediottaviano litemagazines localcommunityhero lustrsazan mariashebets menswheels newwebstechs1 nikkit ootheday oxigenvpn plurfitness prefetial probudgetcarpetcleaning psychicmrsmartin pyaten qvantor rejoignezlogis riselayouts ruchin rwxqsgd5 sakaltime shibiaowang shop-bot skrtips spotlesshaven strcoding tetapatrium topselller uneplanteartificielle vipertechcr webtech1223411 wo-kygaming yamaha-mekar yeknava yoybuysheets zenunwa yamaha-mekar jilidskripsi ankertoto ankertoto ankertoto