ChatGPT Custom Instructions Limit: 1,500 to 5,000 Characters

By Munir Afridi · Updated October 2026 · 12 min read

Custom instruction limits compared (October 2026)

Every major AI assistant now has a place to store standing instructions, and every one of them caps that field differently. The numbers below come from official help pages and the platforms' own builder interfaces. The ChatGPT, Assistants API and Claude figures were rechecked in October 2026.

Two things jump out of that table. First, even inside ChatGPT the cap depends on your plan: 1,500 characters on Free and Go versus 5,000 on paid plans, and 8,000 for a Custom GPT. Second, the platform with the loosest limit (Claude) is also the one telling users most loudly to keep instructions short. Limits and best practice are different things, and this guide covers both.

What is the ChatGPT custom instructions character limit?

ChatGPT custom instructions hold up to 1,500 characters on Free and Go and up to 5,000 characters on Plus, Pro, Business, Enterprise and Education, and the editor will not save text that runs over. The settings cover what ChatGPT should know about you (your role, your industry, your projects) and how it should respond (tone, format, things to avoid). That is roughly 225 words on the lower cap and 750 words on the higher one.

For years the cap was 1,500 characters for everyone. OpenAI has since raised it to 5,000 characters on paid plans, while Free and Go stay at 1,500. The newer personalization screen also added short fields for a nickname and occupation.

What happens if you paste 2,400 characters into a 1,500-character limit? The text gets cut at the cap, usually mid-sentence, and people routinely save the truncated version without noticing. That is the most common reason custom instructions "stop working": the second half never made it in. Count before you paste. Our character counter shows a live count as you edit, so you can trim to 1,495 and keep a small buffer.

How long can Custom GPT instructions be?

8,000 characters. The Instructions field in the Custom GPT builder simply refuses to save anything longer, a wall confirmed by years of OpenAI Developer Community threads asking for more room. That is about 1,200 to 1,400 words of instructions, which sounds like plenty until you try to encode a full brand voice guide, a workflow, and twenty edge cases into one box.

The standard workaround is knowledge files. Builders move long reference material (style guides, product catalogs, prompt libraries) into uploaded files and keep only behavioral rules in the instructions field: what the GPT is, how it answers, and when to consult which file. The instructions become a table of contents instead of an encyclopedia. One popular pattern even adds a self-protective rule: never let the instructions field exceed 8,000 characters while iterating, because the builder will silently block your save and you can lose an editing session to it.

Developers take a different route. The Assistants API was sunset on August 26, 2026, and OpenAI now points developers to the Responses API, where instructions are sent with each request or stored as a reusable prompt. OpenAI's migration guide does not state a fixed instruction-length cap, so check the current API docs before building around a number. The 8,000 figure is a product decision about consumer UX, not a model limitation. If your project genuinely needs a 30-page system prompt, the API is the route. For everyone else, 8,000 disciplined characters beat 50,000 rambling ones.

What is the Gemini Gems instruction limit?

Gemini Gems, Google's equivalent of Custom GPTs, accept roughly 4,000 characters of instructions. Google's own guidance suggests 500 to 2,000 characters is the sweet spot, which matches what Gem builders report: short, structured instructions produce more consistent Gems than maxed-out ones.

Gemini splits personalization across two features, and it pays to use the right one. Gems carry task-specific instructions (a proofreading Gem, a meal-planning Gem). Saved Info holds standing personal facts as short individual entries, things like your dietary restrictions or that you write in British English. Users on Google's support forums have hit walls with both the entry length and the number of entries, so treat Saved Info as a fact list, not an essay field.

Do Claude Projects have an instruction limit?

Not a published one. Claude Project instructions count against the conversation's context window, which is 1M tokens for Fable 5.1, Opus 5.5 and Sonnet 5.5 on paid plans (200K for some other models, such as Haiku 4.5). At an average of 4 characters per token, the theoretical ceiling on the 1M models is around 4 million characters. Nobody should write project instructions anywhere near that, and Anthropic's help docs say so directly: keep instructions concise and focused on essential context, guidelines, and Claude's role.

The reason a giant instruction block backfires is mechanical. Instructions are loaded into every conversation in that project, so every character you add is reprocessed with every message and crowds out room for your actual documents and chat history. Claude's project knowledge handles bulk material more cleverly: on paid plans it switches to retrieval once uploads grow, searching your files and pulling in only relevant passages, so the knowledge base can grow beyond what fits in context. The division of labor is clear. Rules go in instructions. Reference material goes in knowledge. If you want to know how much fits in a Claude conversation overall, our Claude word limit guide breaks down the numbers per plan.

Why do AI platforms cap instruction length?

Three reasons, in descending order of honesty: cost, quality, and simplicity.

Cost is the bluntest. Custom instructions are invisibly prepended to every message you send. A 3,000-character instruction block is roughly 750 tokens, and ChatGPT serves hundreds of millions of users. Multiply 750 tokens by every message from every user who fills the fields and the compute bill is real money. Capping the field caps the bill. You can see exactly what any block of text costs in tokens with our AI prompt word counter, which counts across ten current models.

Quality matters more than users expect. Instruction-following research and plain user experience agree: models obey 5 clear rules better than 40 competing ones. Long instruction blocks bury the important rules in noise, and the model starts ignoring some of them, seemingly at random. Google saying "use 500 to 2,000 of your 4,000 characters" is the platform admitting that the limit is already generous.

And characters are not tokens, which trips people up constantly. Field caps are measured in characters because that is what a settings box can count. Models think in tokens, roughly 4 English characters each. So 1,500 characters is only about 375 tokens, a rounding error against a context window of 200,000 tokens or more. The caps exist at the product layer, not the model layer. Our tokens-to-words converter translates between the two units for every major model.

How do you fit more into the character limit?

Compress. Instruction fields are one of the few places where terse, fragmented writing outperforms polished prose. The model does not need grammar, it needs rules. A few techniques that reliably cut 30 to 50 percent:

Delete the politeness. "I would really appreciate it if you could please try to keep responses fairly concise" is 92 characters. "Be concise." is 12. The model treats them identically.

Use imperative fragments. Write rules the way you would write a sticky note: "UK English. No bullet lists unless asked. Cite sources by name. Ask before assuming budget." Four rules, 93 characters.

Cut examples first. Examples are the biggest character hogs. Keep one if a rule is genuinely ambiguous, otherwise trust the rule. If you need many examples, you have outgrown the field; move them to a Custom GPT knowledge file or a Claude Project document.

State exceptions, not repetition. Instead of restating your tone preference for emails, reports, and social posts separately, write the default once and list only what differs: "Default: plain, direct. Exception: LinkedIn posts can be warmer."

Here is a real before-and-after. Before, 212 characters: "I work as a freelance content writer and I would like you to always make sure that the content you write for me is original, engaging, and optimized for search engines whenever that is possible and relevant." After, 78 characters: "Freelance content writer. Output must be original, engaging, SEO-aware." Same instruction, 63 percent smaller. Run your draft through the word counter as you trim and watch the character count fall without losing a single rule.

Do longer custom instructions actually work better?

Mostly no. There is a point, usually somewhere between 500 and 1,500 characters, where each added rule starts costing more than it returns. Past it, instructions begin to contradict each other, the model averages them out, and you get blander output than you had with no instructions at all. The heaviest Custom GPT builders tend to converge on the same shape: a short identity statement, 5 to 10 hard rules, and pointers to knowledge files for everything else.

This holds across the current model generation. GPT-5.6, Claude Opus 5.5, Claude Sonnet 5.5, and Gemini's 2026 models all follow concise instruction blocks noticeably better than sprawling ones. Newer models are better at long instructions than older ones were, but "better at tolerating bloat" is not a reason to bloat. The 1,500-character cap that feels stingy is, in practice, close to the length where instructions work best anyway.

So treat the limits as a forcing function. The platforms are nudging you toward the instruction length that produces the best output, and the occasional afternoon spent compressing your rules pays off every single message afterward.

How do these fields count characters?

Every limit on this page counts characters, not words, and the counting is stricter than people assume. Spaces count. Line breaks count, and a blank line between paragraphs costs two characters. Punctuation counts. Emoji typically count as two or more characters because they are stored as multi-byte sequences. A neatly formatted instruction block with headers, blank lines, and dashes can spend 10 percent of its budget on pure formatting.

That has a practical consequence: format for density, not beauty. Single line breaks between rules instead of blank lines. Periods instead of decorative dashes. Skip the emoji section markers entirely; the model gains nothing from them and they are the most expensive characters in the field. A character count with spaces is the number that matches what the platform sees, so always check that figure rather than the word count.

If you regularly need long standing instructions, pick the platform accordingly. Claude Projects are the roomiest consumer option since instructions only compete with context space. Custom GPTs sit in the middle at 8,000 characters plus unlimited knowledge files. Plain ChatGPT custom instructions and Gemini Gems are for lean rule sets. And for anything truly industrial, the Responses API, where you control the full system prompt, is the way in. Our AI writing tools hub compares the full lineup, model by model.

Frequently asked questions

Sources: OpenAI Help Center custom instructions page (June 2026); OpenAI Developer Community threads on the 8,000-character GPT builder cap; OpenAI Assistants API documentation; Google Gemini guidance and support forums on Gems and Saved Info; Anthropic Claude Help Center on context windows and project knowledge. Limits change; figures verified June 2026.