Custom GPT Builder: Instructions, Knowledge, and Testing

The Custom GPT builder is the editor inside ChatGPT where you give an assistant a name, a set of standing instructions, and optional knowledge files, then publish it for your own use or for others. The result is a reusable assistant that starts every conversation knowing your rules, your terminology, and the reference material you chose.

A tool like fenbs keeps the approved instructions and the lessons you learn from testing outside of any one assistant configuration, so you can apply them to Claude, Cursor, or any other connected app when you need to.

What the builder gives you

The builder has four main sections: Configure, Knowledge, Capabilities, and Preview. Configure is where you write the name, description, and instructions. Knowledge is where you upload files the assistant can search. Capabilities is where you switch code interpretation, image generation, and web search on or off. Preview lets you test the assistant in a side panel before saving.

OpenAI has announced a transition away from custom GPTs. Check OpenAI's help center↗ (opens in a new tab) for the current availability of the builder on each plan and the timeline for changes.

Writing instructions that hold

Instructions are the most important part of the builder. They run before every conversation and shape every response. A few rules that improve reliability:

• State the single main job first: what this assistant does and what it does not do.
• Write rules as specific commands, not general preferences. "Reply in no more than three bullet points" beats "keep replies short."
• Add a worked example for any rule you have seen broken in testing. The assistant follows examples more reliably than abstract rules.
• Keep the total length as short as it can be while covering every case you actually need. Rules near the bottom of a long list receive less attention.
• When two rules might conflict, say which one wins.

Adding knowledge files

Knowledge files let the assistant search a body of text when answering questions. Plain text, PDF, Word documents, and spreadsheets work well. Very large files, scanned images, and heavily formatted PDFs work poorly.

Before uploading, check each file for statements that contradict your instructions. If a document describes a workflow you have changed, update it first. Outdated content in a knowledge file can silently override what your instructions say. For the current file size and count limits, check OpenAI's documentation↗ (opens in a new tab) directly, as these change with plan updates.

Save the current version of each knowledge file somewhere outside ChatGPT. When you update the assistant, you will need to re-upload, and losing the source means starting over.

Choosing capabilities

Turn on only the capabilities the assistant actually needs. Code interpreter lets it write and run code on uploaded data. Image generation lets it create images. Web search lets it look up current information. Extra capabilities add scope and cost without benefit if the use case does not need them.

If your assistant is meant to answer from its knowledge files and your instructions only, consider turning web search off. An assistant that can search the web may prefer a web result to a carefully curated knowledge file.

Testing before you publish

Use the Preview panel in the builder for an initial check, then open a fresh conversation with the published assistant to test properly. The Preview panel can behave slightly differently from a real conversation.

Write out the five or six inputs you expect most often, send each one, and check whether the output matches your instructions. If something breaks, identify which rule is being ignored before changing anything. Changing multiple things at once makes it impossible to know what fixed the problem.

When a custom GPT is not the right tool

A custom GPT is a good fit for a single repeatable task with stable instructions and a clear audience. It is less useful when the task changes often, when the instructions are long and complex, or when you need the assistant to have access to live project context. For a comparison with other reusable assistant options, see custom GPT alternatives.

The builder also does not store your project state. If the assistant needs to know where a project stands, what decisions have been made, or what happened in previous sessions, that context has to come from somewhere outside ChatGPT. fenbs does not connect to Microsoft Teams or integrate with project management tools automatically, but it does give connected assistants a shared record they can read at the start of each session.

How to do this with fenbs

1. Create a free account and add a project for the work this assistant supports.
2. Add a Skill to fenbs with the full text of your instructions and a short note on what each rule is for.
3. After each round of testing, add a Lessons learned entry: what broke, what you tried, and what fixed it.
4. When you update the GPT's instructions, update the Skill in fenbs so the approved version is always current.
5. Connect Claude, Cursor, or another assistant to the same fenbs project, and it will read the same instructions via MCP without requiring a separate custom GPT.

For related topics, see Claude Projects vs. custom GPTs for a side-by-side comparison, or ChatGPT Projects vs. custom GPTs for the differences within the ChatGPT ecosystem. The OpenAI agent builder covers building more capable agents that can take actions.