Custom GPT Instructions Character Limit: What to Keep

Custom GPT instructions have a character limit, and when you are close to it you have to choose what to keep. The limit forces a useful question: which rules actually need to be in the instructions field, and which can live somewhere else?

The answer almost always splits into two categories: core rules that must shape every reply, and reference material or context that only matters for specific cases. Core rules belong in the instructions. Reference material belongs in uploaded knowledge files or a separate source. Tools like fenbs can hold the maintained prompt alongside the supporting work context, so the instructions field stays focused and the rest is still findable.

Check the current limit

OpenAI sets the limit for custom GPT instructions and may change it over time. Before restructuring your instructions, check OpenAI's current documentation or the editor itself for the exact limit. The editor typically shows a character count or a warning when you are close.

Paste your current instructions into a character-counting tool to see where you stand. If you are well under the limit, you have room to add examples and counter-examples. If you are near it, the sections below explain what to cut.

Separate core rules from reference material

Core rules are instructions that need to apply to every reply, regardless of what the user asks. They include tone rules, scope restrictions, output format requirements, and what to do when the user asks for something outside the GPT's purpose.

Reference material is information the GPT uses to answer specific questions: a product catalog, a list of company names, a set of FAQ answers, or a long style guide. This material matters for certain inputs but not all of them. It does not need to be in the instructions field.

• Keep in instructions: persona, tone, format, scope rules, what to refuse, and one worked example.
• Move to knowledge files: product details, pricing structure, FAQs, long glossaries, style guides, reference tables.
• Move to context when needed: project-specific facts that only apply to one workflow.

Move reference material to uploaded files

Custom GPTs can use knowledge files: documents you upload to the GPT editor that the model can search when answering. A long FAQ document, a product catalog, or a style guide can live in a knowledge file and the GPT will consult it when relevant.

Mention in the instructions that the GPT should refer to the knowledge file when asked about those topics. For example: "For product details, refer to the knowledge file. Do not answer from memory." This keeps the instructions short and ensures the GPT uses the current, accurate source instead of inferring from training data.

Cut explanations, keep rules

Instructions that explain why a rule exists take up space without helping the model. "Be concise because our users are busy professionals who read on mobile" is half as useful as "keep every reply under 100 words." The model applies the rule; it does not benefit from the rationale inside the instructions field.

Save the rationale somewhere you can read it when you revisit the instructions months later. But cut it from the instructions themselves. Every sentence you remove is space for a rule that actually shapes the output.

Prioritize by what you have seen break

If you must cut rules to fit the limit, start with rules you have never seen broken. A rule the GPT already follows without being told is not doing any work. A rule it broke last week is. Prioritize the rules that correct real, observed problems.

Test after each cut. If removing a rule changes the output in an unacceptable way, put it back and find something else to cut. If the output is unchanged, the rule was not needed.

Write a leaner version first

A useful exercise before editing: rewrite your instructions from scratch, including only rules you have seen the GPT break. Start from zero, not from the current long version. Most teams find that the lean version is shorter than the limit and performs as well as or better than the padded one. Rules that come from "this seems like a good idea" rather than "I saw the GPT do this wrong" are often the first to be cut.

How to do this with fenbs

1. Create a free account and add a project for your GPT workflow.
2. Save the full instruction set as a Skill in fenbs, with rationale notes and test examples that do not fit in the instructions field.
3. Move reference material out of the instructions into uploaded knowledge files and note in the Skill which file covers which topic.
4. When you connect Claude, ChatGPT or another supported assistant to fenbs, it can read the full context from the Skill rather than relying on a truncated instructions field.
5. When the instructions change, update the Skill so the complete version is always current.

For related problems, ChatGPT also has a character limit for its personal custom instructions feature — check OpenAI's documentation for the current limit. For instructions that are within the limit but still not working, see custom GPT not following instructions.