QWE AI Academy The biggest mistake people make with ChatGPT custom instructions isn’t writing too little. It’s writing a wish list. “Be concise. Be detailed. Be friendly. Be blunt. Don’t apologize. Use examples. Keep responses short. Explain your reasoning.” Then they wonder why the output feels random.
Contradictions cancel each other out. And since the interface added a Personality dropdown on top of the instructions box, there’s now a second layer quietly fighting whatever you typed. If your setup feels inconsistent, it’s almost never the model. It’s the rules you gave it.
Why the standard tutorial advice falls short
Most guides walk you through the same path: open Settings, find the two boxes, fill in “I’m a marketer who likes bullet points,” save, done. That worked in 2023. It doesn’t anymore.
The current ChatGPT interface stacks three personalization layers that interact with each other: Personality presets, Custom Instructions, and Memory. OpenAI’s personality documentation puts it plainly: if a saved memory contains guidance that conflicts with a personality’s style, it may override or reduce the visible traits of that personality. Translation: your instructions can lose silently to settings you forgot you turned on.
Tutorials that copy-paste the same teacher/developer examples from OpenAI’s original 2023 announcement? Those examples were written before personalities existed. They aren’t wrong – they’re incomplete.
The right setup order: layers before content
Before you write a single word in the instruction box, decide what each layer is for. The principle: lower layers set tone, upper layers set rules.
| Layer | What it controls | When to use |
|---|---|---|
| Personality preset | Default voice and warmth | Pick one and leave it. Treat it as your baseline mood. |
| Custom Instructions | Hard rules: format, length, what to skip | Anything you want enforced every conversation |
| Memory | Facts the model picks up over time | Let it accumulate, but audit monthly |
| Project instructions | Scoped rules for one workstream | When a rule only applies to one body of work |
As of mid-2025, ChatGPT’s personality presets include Default, Friendly, Efficient, Professional, Candid, Quirky, Cynical, and Nerdy – though this list may expand. Pick one that matches the tone you actually want, and your Custom Instructions don’t need to fight it. If you want crisp and businesslike, choose Efficient or Professional; don’t choose Friendly and then write “be blunt” in the instruction box.
How to set up ChatGPT custom instructions step by step
The actual mechanics are short. In Settings, select Customize ChatGPT (or Personalization on some platforms), make sure Enable customization is toggled on, and enter your instructions. Custom instructions are available on all plans – Web, Desktop, iOS, and Android – as confirmed in OpenAI’s help center documentation (as of mid-2025; check there if plan availability has changed).
- Pick your Personality first. Settings → Personalization → Personality. Choose one. Save.
- Open Custom Instructions. Same Personalization screen. Toggle Enable customization on.
- Write constraints, not descriptions. Skip “I’m a designer” – that’s context the model rarely uses. Write “Default response length: under 200 words. Override with ‘go long’ when I want more.”
- Test in a fresh chat. Updates apply immediately to all chats – but old conversations already contain old behavior, which muddies your test.
- Wait a week. Then audit Memory. Check Settings → Personalization → Memory. Delete any saved fact that contradicts your instructions.
A constraint-first instruction template that actually works
Most templates ask you to describe yourself. That produces vague output. A better structure: tell the model what to not do, what format to use, and how to break its defaults.
# Role context (one line, only if relevant)
I work in [field]. My outputs go to [audience].
# Hard rules
- Default reply: under 200 words. If I say "go deep", remove the cap.
- No disclaimers ("As an AI...", "I'm not a doctor...").
- No filler openers ("Great question!", "Certainly!").
- When I share code, return only the changed lines plus 2 lines of context.
- If I ask for an opinion, give one. Don't list pros and cons unless I ask.
# Format defaults
- Prose by default. Lists only when there are 4+ parallel items.
- Bold the conclusion at the top, then explain.
# When uncertain
- Say "I don't know" rather than guessing.
- Flag assumptions explicitly: "Assuming X..."
Each line is a rule the model can either obey or violate – which makes it testable. Vague aspirations like “be helpful and insightful” aren’t testable, which is why they get ignored.
Pro tip: When telling ChatGPT to avoid a behavior, also tell it what to do instead. “Don’t use em dashes” gets weaker compliance than “Don’t use em dashes – start a new sentence or use a comma.” Negative instructions without a replacement leave the model guessing.
The edge cases nobody mentions
Three things break setups in ways the standard tutorials don’t warn you about.
1. Memory drift. Memory accumulates silently. Six months in, ChatGPT “knows” you prefer formal language because of one project from March – and that saved fact is now overriding the casual tone you wrote into Custom Instructions. OpenAI’s documentation is explicit about this: memory can override personality traits entirely. Audit Memory the same way you’d clean a Downloads folder – not when something breaks, but on a schedule.
2. Model-version compliance differs. Per OpenAI’s release commentary (as of mid-2025, and subject to change with future model updates), GPT-5.1 follows Custom Instructions more tightly than GPT-5, so rules like “always answer like X / never do Y” stick better across chats. Instructions that felt like a gentle nudge six months ago may now read as rigid commands. If that’s happening, soften the wording.
3. Custom Instructions don’t add knowledge. They shape behavior, not information. Writing “You know our company’s pricing” doesn’t make it true. Custom Instructions are not retrieval, fine-tuning, or a document knowledge base – they shape how ChatGPT responds; they do not give it access to your private product docs, policy library, help center, or customer records. For source-grounded answers, use Projects with uploaded files or a Custom GPT.
Here’s a question worth sitting with: as Memory gets smarter and accumulates more context over time, where does the boundary between “what you instructed” and “what the model learned” actually sit? Right now you can audit and delete Memory entries. Whether that level of control persists as the feature matures is genuinely unclear.
A real example: rewriting a vague instruction set
Here’s a typical “wish list” someone might paste in:
I'm a startup founder. I want detailed strategic advice but also
quick answers. Be professional but friendly. Don't be too verbose
but explain your reasoning. Help me think through problems.
Four contradictions in five sentences. Detailed but quick. Professional but friendly. Not verbose but explanatory. The model picks whichever pattern its training nudges it toward – and you blame ChatGPT for being inconsistent.
The rewrite:
Default mode: 3-sentence answer + one follow-up question.
If I type "deep dive", switch to long-form with reasoning shown.
If I type "decide", give one recommendation and the trade-off, nothing else.
Skip pleasantries. Open with the answer, not the framing.
Now there are three named modes the model can switch between, and one explicit default. No contradictions to resolve.
Privacy and the part most guides skip
Turns out there’s a data angle most setup guides don’t mention. OpenAI’s original custom instructions announcement notes that information from your use of custom instructions may be used to improve model performance – unless you opt out via data controls. Don’t paste API keys, client names, unreleased product details, or anything you’d be unhappy seeing surface in a future model’s training set. Use placeholders: “Client A,” “Project Zeta.” Same behavior shaping, without the exposure.
FAQ
Do I need ChatGPT Plus to use custom instructions?
No. Free on every plan.
Why does ChatGPT keep ignoring my custom instructions?
Start with the most common culprit: contradictions. “Be brief” and “explain thoroughly” can’t both be true – the model picks one and you get inconsistency. Check for that first.
If the instructions look clean, the next place to look is the Personality dropdown. A “be direct” rule written into the instruction box loses to the Friendly preset on tone – the preset sits at a lower layer and shapes the baseline voice before your rules even apply. Swap the preset to Efficient or Professional and test again in a fresh chat.
Still broken? Open Settings → Personalization → Memory. An old saved fact – something from a project months ago – may be overriding your current rule. Delete it. Fresh chat. That usually fixes it.
What’s the difference between Custom Instructions, Memory, and Projects?
Custom Instructions are your global rules. They apply to every chat, every time, until you change them.
Memory is different in kind, not just scope – it’s what ChatGPT picks up from actual conversations over time. You didn’t write it; the model inferred it. That’s why it can silently override your instructions if the two conflict. You can edit or wipe Memory directly.
Projects let you scope rules to one body of work: “all chats in this Project output in Markdown with H2 headers.” If a rule only matters for one workflow, keep it out of your global instructions – it’ll just create noise everywhere else.
Open Settings → Personalization right now and audit what’s actually in there. If the box is empty, fill it with the constraint template above. If it’s full of vague descriptions, rewrite one rule into a testable constraint and watch what changes in your next conversation.