Why Is My Prompt "Too Long"?

TL;DR: "Prompt too long" is three unrelated problems wearing one error message. It might be a real character cap on a specific field (ChatGPT custom instructions: 1,500 or 5,000 characters). It might be the conversation's whole context window filling up, which has nothing to do with your last message. Or there's no limit at all, and a diluted, contradictory prompt is just performing worse.

You type something that used to work, or something that looks perfectly reasonable, and the tool balks: a rejected field, a reply that cuts off mid-sentence, or an answer that quietly gets worse the longer your message gets. "Too long" names the symptom, not the cause, and the three real causes call for three different fixes. Fix the wrong one and you'll trim words that were never the problem, or worse, cut the one sentence that was actually doing the work.

This post separates them: the character limits that are real and published, so you stop guessing at a number nobody wrote down; the context window, a conversation problem wearing a prompt-length costume; and the case with no limit in sight, where the fix is a cutting method, not a word count, with more than twenty before/after pairs you can apply directly.

Is There Actually a Character Limit, or Does It Just Feel Long?

Each of these has a different fix, so work through them in order.

1. A hard limit. A specific field on a specific surface rejects input past a stated character count. Verifiable, and there are exactly two numbers OpenAI publishes for this.
2. The context window filling up. Not about your prompt at all; it's about the whole conversation, including every earlier turn and any attached files.
3. No limit at all. Your prompt is fine by every measure, and it still performs worse than a shorter version because it's diluted.

Rule out the first one fast, since it's the only one with an actual number attached to it.

Where Are the Real, Published Character Limits?

One surface, two numbers. OpenAI's help centre states plainly: "Free and Go users can save up to 1,500 characters in custom instructions." It adds: "Plus, Pro, Enterprise, Business, and Education users can save up to 5,000 characters." (help.openai.com, verified 2 September 2026.)

That's the only per-field prompt-length cap OpenAI states in plain numbers, and it applies specifically to the custom-instructions field in ChatGPT settings, not to a normal chat message. We could not find a published character limit for a single chat message anywhere in OpenAI's own docs (checked 2 September 2026). If a very long message causes a problem, it's unlikely to be a stated cap. See the context-window section below.

What About Project Instructions and Custom GPTs?

Not published, on either surface, and worth saying plainly rather than guessing.

Anthropic's Claude support articles describe plenty about Projects (what counts as project knowledge, how retrieval works, what free and paid accounts get) without ever stating a character or token limit for the project instructions field itself, across the articles we checked (support.claude.com, 2 September 2026). If you find a specific number quoted for this online, it isn't Anthropic's.

Custom GPTs are murkier for a different reason. OpenAI's help centre states: "New GPT creation and publishing are not available on personal ChatGPT accounts, including Free, Go, Plus, and Pro." (help.openai.com, verified 2 September 2026.) Existing GPTs remain usable and editable where plan and permission requirements allow it. OpenAI has never published a Custom GPT instruction-length limit, and on a personal account you likely can't create a new one to test the question anyway.

Is It Actually the Context Window Filling Up?

A completely different problem, and it has nothing to do with how long your last message was.

The context window is the fixed amount of a conversation a model can hold at once: your messages, the model's replies, and anything you've attached or pasted in, all competing for the same space. You can write a short, careful prompt and still run into trouble forty messages into a long chat, because it isn't that message that's full. It's everything before it.

What this looks like varies by platform and isn't consistent enough to name one symptom. Sometimes it's an explicit error. More often it's quiet: the model stops referencing something you said early on, contradicts a decision from ten messages ago, or a reply cuts off mid-sentence with no warning, a symptom we cover on its own in Why Does ChatGPT Cut Off Mid-Answer?. Either way, the fix isn't trimming your latest prompt. It's managing the conversation: starting a fresh chat when the old one has drifted, covered in When to Start a New Chat vs Keep Going, or reaching for a tool built around long documents, like Kimi K3 Prompt Templates (Long Context).

We're not going to hand you a context-window size here. Every figure we could quote would be stale within weeks, tied to a specific plan, or, in at least one case we checked while researching this piece, labelled by the vendor itself as illustrative rather than real. Check the number on the platform you're actually using, on the day you're using it.

What If There's No Limit at All, and It Just Got Worse?

This is the case worth understanding properly, because there's no error to point at. Your prompt is under every limit that exists, and it still produces a worse answer than a shorter version of the same request would.

OpenAI's own prompting guidance for GPT-5.6 opens with this: "Removing repeated instructions and examples and simplifying tool descriptions can improve task performance and token efficiency" (accessed 2 September 2026). It gets more specific a few lines later, on the part that actually costs you: "GPT-5-class models follow prompt contracts closely, so conflicting rules can create more instability than missing detail." A gap in your prompt gets filled with a reasonable default. A contradiction gets resolved by a rule you didn't choose.

The same mechanism shows up on a completely different surface. Anthropic's guidance for Claude Code, about keeping a project's standing instruction file short, doesn't hedge: "Bloated CLAUDE.md files cause Claude to ignore your actual instructions!" (code.claude.com, accessed 2 September 2026). That's advice about a file loaded once per session, not a chat prompt, but the underlying problem is the same one: attention is scarce, and volume competes against it regardless of which surface is carrying the volume.

We work through the deeper argument for why length itself isn't the lever, with vendor sources on both sides, in Does a Longer Prompt Get Better Results? (Density vs Padding). Here, we only need the practical output: what to cut, and in what order.

What Should You Cut First?

Roughly in this order, since each pass makes the next one easier to see:

1. Restated rules. Anything you've said more than once, in any words.
2. Politeness and softeners. Greetings, apologies, "if it's not too much trouble."
3. Stakes and hype. "This is very important," tip offers, all-caps emphasis.
4. Duplicate examples. Two examples that illustrate the exact same shape.
5. Contradictions. Read what's left, specifically for two instructions that can't both be followed, and pick one rather than leaving both in.

Here are more than twenty of these cuts, pulled from prompts people actually send, sorted by type.

One full pass, side by side, on a support-response prompt:

BEFORE (127 words, mostly packaging)
Hi! I hope you're having a great day. I really need your
help with something important, so please take this
seriously and do your best. Please write a response to this
frustrated customer. Make sure the tone is professional —
this is very important, keep it professional throughout the
whole reply. Be thorough and detailed and cover everything.
Also, please double-check that you've addressed their
concern properly. I know this is a lot to ask, but if you
could take extra care with this one, that would mean so
much to me. Thanks in advance, I really appreciate it! One
more thing — try to keep it fairly short if you can.

AFTER (41 words, same instructions, no contradiction)
Reply to this frustrated customer about a delayed refund.

Tone: professional, empathetic.
Length: under 120 words.
Must include: the refund will post within 3 business days,
and a direct apology for the delay.
Do not: offer a discount or promise a specific dollar amount.

The "before" also buried a real contradiction: "cover everything... be thorough" against "keep it fairly short if you can," left for the model to resolve however it liked. The "after" version resolves it once, on purpose.

What Should You Always Keep?

The specifics only you have. A model can generate plausible-sounding filler about almost any topic; it cannot invent your audience, your exact constraints, or the one detail that makes this request different from a generic version of itself.

That means keeping: who's reading the output, the exact format it needs to land in, the hard constraints that actually bind (a word count that matters, a word that must never appear, a fact that must stay accurate), real source material instead of a description of it, and what "done" looks like well enough that the model knows when to stop. None of that is padding, no matter how long it makes the prompt. For a broader pre-send checklist that catches more than length, see The Prompt Hygiene Checklist.

How Do You Count This Honestly?

Not in words, and be careful with round numbers you've seen floating around.

A token is not a word. Common English words often land close to one token each, but punctuation, uncommon words, code and non-English text routinely split into more tokens than their word count suggests. Treat character and word counts as rough proxies, useful for a quick gut check, useless for anything precise. If a platform actually shows you a token count, use that instead of estimating.

Where Does a Prompt Tool Fit Into This?

On the "match the mode to the actual problem" side, not on "always make it shorter."

Prompt Architects ships four enhancement modes because "too long" and "too vague" are different failure states that need opposite fixes. Refine adds structure to a prompt that's missing a role, a format, or a constraint. Shorten does the reverse: it strips a bloated prompt down to its essential instructions without dropping the ones that still matter, and it's available starting on the Pro plan ($4.99/month at the time of writing) per our pricing page. The Free plan, which includes 5 prompt enhancements a day forever per our FAQ, doesn't include it.

Neither mode makes a prompt better by making it longer for its own sake. The before/after view exists so you can see whether a rewrite added a real constraint or just padding, and cut it back yourself if it's the latter: the same subtraction the table above walks through by hand.