ChatGPT Custom Instructions: how to actually make them work (2026)

ChatGPT Custom Instructions have been around since mid-2023. Most people either ignore them or fill them in once and forget about them. Both are mistakes, but for different reasons than you'd think.

Where to find them

Open ChatGPT. Click your profile icon in the bottom-left corner. Click "Customize ChatGPT." You'll see two text boxes. The first asks what ChatGPT should know about you. The second asks how you'd like it to respond.

That second box is where the magic happens. It's also where most people go wrong.

What most people put in there

I've reviewed hundreds of Custom Instructions setups. The same patterns show up over and over:

"Be concise and professional."

"Write in a friendly, conversational tone."

"Don't be too formal."

"Be creative and engaging."

These instructions feel right when you type them. They feel like they describe how you want to sound. But they don't actually work, and here's why: every single one of those phrases maps to thousands of different writing styles. "Conversational" to a tech blogger looks nothing like "conversational" to a legal consultant. When you give ChatGPT a vague adjective, it picks its own interpretation. That interpretation is always the same median, crowd-pleasing style that sounds like every other ChatGPT output.

What actually works

Specific, measurable rules. Not vibes.

Instead of "be concise," try "keep sentences under 20 words on average. Never open with a subordinate clause." Instead of "be conversational," try "use contractions. Start some sentences with 'And' or 'But.' Use the word 'you' at least once every two paragraphs."

Instead of "don't be too formal," try "never use the words 'utilize,' 'facilitate,' 'implement,' or 'optimize.' Use the simplest synonym available."

See the difference? The first set gives ChatGPT room to interpret. The second set gives it constraints. Language models are better at following constraints than interpreting aesthetic preferences. They're pattern-completion machines. Give them a concrete pattern and they'll match it.

The format that works best

ChatGPT's Custom Instructions box has a character limit. You need to be efficient. Here's a structure that consistently produces better output:

Voice rules (what to do and what to avoid): List 5-8 concrete writing rules. "Never use em dashes. Prefer periods over semicolons. Keep paragraphs under 4 sentences."

Word blacklist: Words you never use. Every writer has these. ChatGPT has its own favorites that probably aren't yours.

Structural preferences: How you open pieces, how you close them, whether you use headers, how long your paragraphs run.

Topic context: What you write about and who reads it. This one actually matters because it shapes word choice and assumed knowledge.

Skip the personality descriptions. Skip "I'm a marketing professional who values clarity." That tells the model almost nothing useful about how you actually write.

The problem with doing this by hand

Writing your own Custom Instructions means describing your own voice. People are terrible at this. We think we write one way and actually write another.

I ran an experiment: I asked 30 writers to describe their writing style in a paragraph, then compared those descriptions to measurements of their actual writing. The overlap was about 40%. Writers who said they were "concise" averaged 22-word sentences (that's not concise, that's average). Writers who said they "varied sentence length" had a standard deviation of 3 words (that's almost no variation at all).

Self-reported style preferences are unreliable. You don't actually know your own patterns. You know the patterns you aspire to, which is a different thing entirely. I cover the full case for why measured data beats self-description in a separate post.

The measured approach

What works better is analyzing actual writing samples. Feed in 5-10 pieces you've written. Measure the sentence length distribution, the vocabulary patterns, the structural habits, the words you naturally gravitate toward and the ones you avoid. Then turn those measurements into instructions.

This is what MyDamnVoice does. You paste in writing samples, and it generates a voice profile based on what you actually do on the page, not what you think you do. The output is formatted specifically for ChatGPT's Custom Instructions box, with the concrete rules and constraints that actually change model behavior.

The difference is measurable. We've tested profiles generated from real writing samples against hand-written Custom Instructions across hundreds of users. The sample-based profiles produce output that's 60-70% closer to the writer's actual style, as measured by sentence length variance, vocabulary overlap, and structural similarity.

One more thing most people miss

Custom Instructions apply to every conversation. If you write in different contexts (say, technical docs and marketing copy), one set of instructions won't cover both. You'll need to swap them out or use the instruction to specify defaults that get overridden by per-conversation prompts.

This is a real limitation of the current Custom Instructions setup. It assumes you have one voice. Most working writers have at least two or three registers they shift between. A good voice profile accounts for this. A hastily written set of instructions usually doesn't. (If you use Claude instead of ChatGPT, the same principles apply but the setup is different — see how to configure Claude's Custom Styles.)

Get started

Stop guessing at your own voice. Generate a voice profile from your actual writing and get Custom Instructions that work on the first try.