ChatGPT custom instructions are capped by characters, not by words, and the cap depends on what you pay. That one detail breaks most of the advice written about them, because a prompt built for Claude or Gemini is usually several times too long and there is nothing in the interface that explains why it will not hold.

The limits, from OpenAI's own documentation, are 5,000 characters on Plus, Pro, Enterprise, Business and Education, and 1,500 characters on Free and Go.

Why the ChatGPT custom instructions limit catches people out

Characters and words are different units, and the gap between them is larger than intuition suggests. English academic prose averages roughly six characters per word once spaces and punctuation are counted, so 1,500 characters is around 240 words. Not 1,500 words. A standing instruction set of the kind people paste into Claude runs to two thousand words or more, which is somewhere north of 13,000 characters.

That is eight times the Free allowance and nearly three times the paid one. The practical result is that anyone following a guide written for a different tool pastes something, sees it accepted or silently clipped, and assumes the instructions are in force when most of them never arrived.

Truncating is worse than condensing

The instinct when something is too long is to cut the end off. For a writing prompt this is the worst available option, because the end is where the checks live.

A well-built academic prompt is ordered roughly like a process: what to read first, how to write, how to cite, then a final review pass that catches a missing thesis, an uncited claim, or an overstated result. Truncate it and you keep the stylistic preferences and lose the verification. You end up with output that has the right tone and none of the discipline, which is the combination hardest to spot when you are reading your own draft.

Condensing keeps every rule and shortens how each one is expressed. It is more work than trimming, and it is the difference between a prompt that shapes a paper and a prompt that only changes its accent.

What survives the cut

When something genuinely has to go, the order below reflects how much each part changes a finished draft.

  • Keep the anti-fabrication rule. Never invent facts, quotations, statistics or citations, and say so when a source cannot be read. This is the one rule whose absence produces a confident essay resting on references that do not exist.
  • Keep PEEL. Point, evidence, explanation, link. Models stack claims and citations and skip the explanation, and that missing step is the clearest difference between a marked essay and raw output.
  • Keep the punctuation ban. No em dashes, colons or semicolons. Cheap to state, and it removes the most recognisable surface marker in one line.
  • Keep evidence-tied hedging. Cautious phrasing where evidence only suggests, direct phrasing where it is established.
  • Cut the worked examples first. Lists of preferred transitions and sample citation phrasings eat characters faster than anything else and contribute least per character spent.

Where the box is

ChatGPT custom instructions live in one place. Open Settings, then Personalization, then Custom instructions. Paste and save. It applies to new conversations rather than retroactively to ones already open, which is worth knowing if you test it in a chat that was already running and conclude it did nothing.

One habit worth adopting whatever your plan: keep the full-length version in a file somewhere. Character limits change, plans change, and the condensed version is lossy. Rebuilding it from memory a month later is harder than it sounds.

What custom instructions will not do

They will not lower an AI detector score in any reliable way. Instructions shape structure, register and vocabulary. Detectors measure how predictable each word is given the ones before it, which is a property of how the model samples text rather than of what it was told to sound like. What detectors actually measure explains the gap, and prompting alone moves a score less than people expect.

Better instructions do make the draft better, which is a separate and more durable benefit. Where evenness survives into the finished text, restructuring it afterwards is a different operation from generating it differently, and it is what an AI to human rewriter does to text you already have. AI Rewriter's human text rewriter handles that half of the job, and the two pair well, since even a well-instructed draft tends to arrive evenly paced.

They also will not make the argument yours. A correctly structured, properly hedged, well-cited essay produced by a model is still produced by a model, and that is the question an academic integrity process asks. Use these instructions on your own reasoning and on sources you have actually read.

Ready-made ChatGPT custom instructions sized for both tiers, alongside the full-length prompt for tools that accept it, are on the setup guide, with the character count shown against each limit.