Personalize ChatGPT Around Your Workflow: What to Put Where

Most people personalize ChatGPT exactly once. They paste a paragraph into custom instructions on a Sunday night, it behaves beautifully for a week, and then one morning it's back to bullet-pointing everything and opening with 'Great question!' The natural reaction is to write the rules louder. That rarely helps.

The problem usually isn't your wording. It's that the right information is sitting in the wrong place, or you're testing somewhere your settings don't even reach. So skip the settings-menu tour. Below is a workflow-first map of what goes where, a checklist for when things drift, and the one habit that keeps your setup portable between ChatGPT, Claude and Gemini.

Why ChatGPT custom instructions seem to stop working

Start with the detail that explains a surprising number of complaints. Per OpenAI's own help text, as quoted by Memorylake, 'GPTs do not use saved memory, custom instructions, or past conversations.' If you built a Custom GPT for work, or someone shared one with you, your carefully tuned account preferences simply aren't in the room. A responder on the OpenAI Community forum put it plainly: a GPT runs on its own Instructions box plus whatever files and tools are switched on, and your account-level Customize settings get overridden.

That's not the whole story, though. Even inside the instructions a GPT does read, people report slippage. In an OpenAI Community thread titled 'Custom GPT not following all instructions', one user wrote: 'I've written instructions for resume generation and the session is not following all the instructions.' Another said it 'routinely ignores specific grammatical instructions every single instance I ask it for something.' A third described a GPT that kept using a banned word even after ChatGPT itself wrote the fix.

And it isn't just an old-model problem. A June 2026 thread reported that on GPT-5.5, 'the first response already ignores or deviates from those instructions' in both Custom GPTs and Projects. OpenAI Support's reply in September offered no fix, only a request to open a support case. A linked thread had the most honest title of the lot: 'Projects are great but Instructions get ignored.'

Run this checklist before you rewrite anything

Memorylake's troubleshooting list is the most useful part of its guide, and its key point runs against instinct: stronger wording and piling on more rules don't fix drift. Check the boring stuff first.

• Is personalization actually switched on? It's a toggle, and it's easy to miss.
• Are you inside a Custom GPT? Then your account instructions don't apply, and the rules need to live in that GPT's own Instructions box.
• Did you give the rule in chat instead of the settings field? A preference typed mid-conversation is just a message, not a standing instruction.
• Are you testing in an old chat? Start a fresh one after every change.
• Are you near the character limit? Blogs quote per-plan caps, but they shift, so trust what your own settings page allows over a number you read somewhere.
• Do your rules fight each other? 'Be concise' and 'explain your reasoning' can't both win, so the model picks one per reply.

What to put where when you personalize ChatGPT

That last checklist item is the sneaky one, because most instruction sets grow by accretion, one annoyed fix at a time, until they contradict themselves. The cure is a division of labor. Each layer is built for a different kind of information, and in our experience drift starts when one layer is quietly doing another's job.

• Custom instructions: stable preferences that apply to nearly everything. Your role, your preferred format, words you never want to see, how much hand-holding you like.
• Memory: context that evolves over months, like your current tool stack or the project you're halfway through. Save recurring context and skip one-off details, or memory fills up with noise.
• Projects: one client, codebase or course, with its own files and instructions. Anything that's only true for that body of work belongs here, not in your global settings.
• Custom GPTs: a repeatable role, like a resume reviewer or a meeting-notes summarizer, that you'd reuse across unrelated projects.

ChatGPT Projects vs Custom GPTs, and the Claude and Gemini versions

A quick gut check: if your global instructions mention a specific client by name, something's in the wrong layer. Move it into a Project and your everyday chats get cleaner as a bonus.

The same split holds across vendors; only the names change. Bootstrap Creative's terminology map lines them up: a Custom GPT or a Gemini Gem for a recurring role, a ChatGPT or Claude Project for an ongoing engagement. Their note on depth matters too. Claude Skills can carry far more detail than a Gem's short instruction field, so a role that needs a long, step-by-step procedure fits better there.

Here's the catch that companies selling memory add-ons don't love to lead with: there's no standard way to export or import instructions or memory between ChatGPT, Claude and Gemini. Whatever you build inside one chatbot stays inside it.

Keep a master context file outside any one chatbot

This is the habit that makes everything else manageable. Write your preferences once, in a plain markdown file you own, and paste the relevant slice into whichever tool you're using. When an instruction set drifts, you're not rebuilding it from memory. You're comparing against the original.

Keep it short enough to read in a minute. When a new pet peeve shows up, add it here first, then sync it out. The prompts you rerun every week deserve the same treatment, saved as templates rather than retyped. Our Prompt Crafter tool is built for that, and the guide to prompt engineering for daily productivity covers how to structure them.

# My AI context (master copy)

## About me
Role, industry, what I usually use AI for.

## Always
- Format: short paragraphs, no intro fluff
- Ask one clarifying question if the task is ambiguous

## Never
- Words or phrases I don't want to see

## Per-project notes
- Client A: tone, audience, files to reference

When the model still ignores your instructions

Sometimes you'll do everything right and it still slips, especially deep into a long chat. The workarounds users trade on the OpenAI forum are low-tech, but they work often enough to be worth a try.

• Ask it to reread its instructions before the next task, mid-conversation.
• Have it restate your instructions in its own words, then check that summary against what you meant. A mismatch tells you which rule is ambiguous.
• Write instructions as ordered steps rather than a pile of preferences. 'First do X, then Y' holds up longer than 'always remember X'.

Personalization isn't a dossier

One last thing worth unlearning. The Slow Journalist's piece on what ChatGPT knows about you notes that memory and custom instructions are genuinely confusing, and that viral 'roast me' prompts mostly extrapolate patterns from your chats rather than reveal some hidden profile. What actually shapes the answers you get is a few layers of text you control. Place each piece deliberately, keep the master copy somewhere you own, and you'll spend a lot less time arguing with your own settings.

References

• OpenAI Community - "GPT-5.5 ignores Custom GPT and Project instructions at the start of new chats"
• OpenAI Community - "Custom GPT not following all instructions"
• Memorylake - "Stop ChatGPT forgetting custom instructions"
• Bootstrap Creative - "ChatGPT vs Claude vs Gemini: Feature & Terminology Map (2026)"
• AI Memory - "ChatGPT Projects vs Memory vs Custom Instructions: Complete Guide (2026)"
• The Slow Journalist - "What does ChatGPT know about me?"