How I Set Up Custom Instructions So I Stop Repeating Myself in Every Chat

For about a year, the first two minutes of every AI conversation were the same. I'd open a new chat, and before I could ask for anything useful, I'd paste the same block of context. Who I am. What I'm working on. How I want it to write. What I don't want it to do. Then I'd ask my actual question.

I did this hundreds of times. Same paragraph, retyped or re-pasted, every session. And half the time I'd forget one of the constraints, get output that ignored it, and have to correct the model mid-conversation — which is the exact correction I'd made yesterday, in a different chat, that vanished the moment I closed the tab.

Custom instructions fixed this. Every major AI tool now has a spot for standing context that gets injected into every conversation automatically — ChatGPT calls it Custom Instructions, Claude has a similar personal preferences field, most tools have some version. You write it once. It applies to every chat forever. It's the single highest-leverage setup step I've done, and most people either skip it or fill it with the wrong things.

Here's what I've learned about what actually belongs there.

The rule: put in what's true every time, not what's true this time

The mistake people make with custom instructions is treating them like a prompt. They stuff in task-specific detail — "help me write my Q3 marketing plan" — which is useless the moment the task changes.

Custom instructions are for the things that are true across *every* conversation you'll ever have. Who you are. What you do. How you want the model to behave by default. The stuff you'd otherwise retype every single time.

The test I use: if I'd have to paste it into more than half my chats, it belongs in custom instructions. If it only applies to one task, it belongs in the prompt.

Part 1: Who you are and what you do

The model produces dramatically better output when it knows the context you're operating in. Not your life story — the specific facts that change how a good answer is shaped.

This is the "about me" block I use:

I'm a solo founder building software and creative products. I work across code, product, writing, and video. I have strong technical background but I'm not a specialist in any one area. When you explain something, assume I can handle depth — don't oversimplify. When you recommend tools or approaches, assume I'm optimizing for shipping fast alone, not for a team. I care about output that's usable as-is, not drafts I have to heavily rework.

Category: setup. Tool: ChatGPT / Claude. Status: permanent. Every sentence here changes the output. "Solo founder" kills every recommendation that assumes I have a team. "Don't oversimplify" stops the model from explaining things I already know. "Usable as-is" raises the bar on every response.

The thing to notice: none of this is about a specific task. It's the frame that makes every task-level answer better.

Part 2: How you want it to write and behave

This is the part that saves me the most re-typing. Every model has default behaviors I don't want — hedging, over-explaining, agreeableness, the same three-part essay structure on everything. I used to correct these one chat at a time. Now I set them once.

This is my behavior block:

Default to direct answers. Skip the preamble — don't restate my question back to me, don't open with "Great question." Lead with the answer, then support it. Don't hedge with "it depends" unless you then explain what it depends on. When I ask for feedback, be critical by default — I'm asking because I want the flaws, not the reassurance. If I'm about to make a mistake, say so plainly. Don't pad responses to seem thorough. A short right answer beats a long complete one.

Category: setup. Tool: ChatGPT / Claude. Status: permanent. The line that does the most work is "skip the preamble." Compounded across every conversation, the amount of throat-clearing this removes is enormous. The critical-feedback line means I no longer have to authorize disagreement in every individual chat — it's on by default, everywhere.

One warning: don't over-constrain the voice here. I keep behavior rules in custom instructions but leave *content voice* (how a specific piece of writing should sound) to the individual prompt, because it changes per task. A blog post and a client email want different voices. "Be direct and skip preamble" applies to both. "Write in a warm, conversational tone" only applies to one.

Part 3: The standing corrections

This is the part almost nobody uses, and it's my favorite. Every time I catch myself making the same correction across multiple chats, I add it to custom instructions so I never have to make it again.

These accumulate over time. Mine currently includes things like:

When you write code, don't add comments explaining obvious lines. When you give me options, give me your recommendation too — don't just list trade-offs and make me decide blind. When I ask you to edit my writing, show me only what you changed and why, not a full rewrite I have to diff manually. Never use the words "delve," "leverage" as a verb, "in today's landscape," or "it's important to note." If you're uncertain about a fact, say so instead of stating it confidently.

Category: setup. Tool: ChatGPT / Claude. Status: permanent — always growing. Each of these lines is a correction I got tired of making. "Give me your recommendation too" came from months of getting neutral option-lists when I wanted a decision. The banned-words list came from noticing the same AI tells over and over. Every line here is a specific annoyance I fixed permanently instead of fighting daily.

The way I maintain this: whenever I find myself typing a correction I've typed before, I stop and move it into custom instructions. The friction of doing it once is trivial. The payoff is that the correction applies retroactively to every future conversation.

What I deliberately keep out

Custom instructions have a length limit, and more importantly, everything you put in there competes for the model's attention on every single request. Bloat them and the important instructions get diluted. So I'm ruthless about what stays out:

• Task-specific context. That's what the prompt is for. My marketing plan doesn't belong in a field that also loads when I'm debugging code.
• Anything I only need occasionally. If it applies to 10% of chats, it goes in a saved prompt I paste when relevant, not in the always-on instructions.
• Long examples. One or two short style cues are fine. A 300-word writing sample in custom instructions is loaded on every chat, including the ones where I'm asking about a recipe. Keep examples in task prompts.
• Contradictions. If two instructions conflict, the model picks one unpredictably. I re-read the whole block occasionally to catch drift.

The mental model: custom instructions are the operating system, not the application. They set defaults. The prompt is where the actual work goes.

The one-time setup that pays forever

Here's what changed when I finally did this properly. My conversations start at the answer now. I don't spend the first exchange establishing who I am and how I want to be spoken to — the model already knows. The corrections I used to make daily are made zero times, because they're baked in. And the output is more consistent, because the baseline is the same in every chat instead of depending on whether I remembered to include a given constraint that day.

If you use AI every day and you've never filled in custom instructions, this is the highest-return hour you can spend. Write the three blocks above — who you are, how you want it to behave, and your standing corrections. Then treat it as living: every time you catch yourself making a repeat correction, move it in.

The whole point of a tool you use daily is that it should get to know you. Custom instructions are the only place that memory actually lives. Everything else resets when you close the tab.

(This is also why I keep my heaviest, most-tuned prompts in Super Prompts rather than in custom instructions — the always-on field is for defaults, and the saved library is for the specific, powerful prompts I pull out when a task calls for them. Different jobs, different homes.)