QWE AI Academy There are two ways to build a GPT custom assistant. The first is to chat with the GPT Builder in plain English and let it figure things out. The second is to skip that conversation and go straight to the Configure tab. Both work. Only one of them produces a GPT that’s still useful three weeks later.
The Configure-first approach wins because it forces you to think about instructions, knowledge files, and capabilities as separate things – which is exactly how the underlying system treats them. The chatty builder blurs all of that together and gives you a vague assistant that sounds smart in preview and forgets your rules in production.
Why most custom GPTs quietly stop working
The typical failure mode isn’t dramatic. You build a GPT, it nails a few demo prompts, you share it with a teammate, and a week later they’re back asking ChatGPT the same questions manually. The GPT didn’t break. It just never had clear enough rules to stay useful under messy real-world prompts.
The root cause is almost always the same: mixing behavior rules into knowledge files, or forgetting to tell the GPT when to read those files at all. OpenAI’s documentation is explicit here – rules, tone, and workflow guidance go in instructions; knowledge is for reference material only. Swapping those two is the fastest way to get a GPT that seems to work and then doesn’t. And uploading a PDF doesn’t mean the GPT will read it. The Zapier team testing a custom GPT found their assistant wasn’t consistently pulling from the uploaded guide – because the instructions never told it when to look there.
The build sequence that actually holds up
Open chatgpt.com/gpts, hit Create, and immediately switch to the Configure tab. You’re going to fill in five fields in a deliberate order – not because the UI demands it, but because each field constrains the next.
- Name and description – write the description first as if you’re explaining the GPT to a new hire in one sentence. The name follows from that.
- Instructions – this is the actual program. Define the role, the format of outputs, what to refuse, and crucially, when to consult knowledge files. Without that last part, your uploads are dead weight.
- Conversation starters – these aren’t decoration. They train users to ask the GPT what it’s good at, which keeps it inside its scope.
- Knowledge – upload reference material only. You can attach up to 20 files, each up to 512 MB (per OpenAI’s help documentation), but bigger isn’t better. Long, messy PDFs tend to underperform short, clean text files.
- Capabilities – turn off what you don’t need. If your GPT writes copy, it doesn’t need Code Interpreter. Every enabled capability is another way for it to wander off-task.
Then test it. OpenAI Academy recommends writing 10 to 15 questions that reflect the tasks your GPT should handle – including known correct answers – then running them to check whether outputs are accurate and adjusting instructions or knowledge where they’re not. This is the step every tutorial mentions in passing and nobody actually does.
A worked example: meeting notes to ticket drafts
Skip the brand voice example every other tutorial reaches for. Try this instead: a GPT that takes raw meeting notes and outputs draft Jira tickets in your team’s format.
The instructions field for this looks roughly like:
You are a ticket-drafting assistant for a product engineering team.
INPUT: pasted meeting notes, often messy.
OUTPUT: one draft ticket per actionable item, in this exact format:
Title: <imperative verb + short noun phrase>
Type: bug | feature | chore
Description: 2-4 sentences, user-facing.
Acceptance criteria: bulleted list, testable.
RULES:
- If the notes mention a deadline, surface it at the top of the ticket.
- If something is ambiguous, mark it [NEEDS CLARIFICATION] instead of guessing.
- Consult the uploaded "ticket-style-guide.md" before formatting. If it conflicts with these instructions, the instructions win.
- Refuse to draft tickets for anything not in the notes.Notice what that does. It tells the GPT exactly when to consult the knowledge file (before formatting), how to handle conflicts (instructions win), and what to do with ambiguity (flag it, don’t fabricate). That’s the difference between a GPT that works and one that quietly invents acceptance criteria.
The gotchas tutorials don’t mention
A few things you’ll discover only after shipping your first GPT:
| Gotcha | What it means |
|---|---|
| GPT-4o retirement | ChatGPT Business, Enterprise, and Edu customers retain access to GPT-4o within Custom GPTs until April 3, 2026, after which GPT-4o is fully retired across all plans (per OpenAI’s help center). Any GPT whose instructions were tuned against GPT-4o’s specific behavior will act differently afterward – re-test before the cutoff. |
| Mobile is read-only | Building and editing GPTs is limited to the web. Mobile apps let you use GPTs but not build them. If you want to tweak instructions from your phone, you can’t. |
| Knowledge file types | Accepted file types vary by model and some are only available if Code Interpreter & Data Analysis is enabled. A spreadsheet upload may not work if Code Interpreter is off – worth checking before you assume the file is available to the GPT. |
| Paid plan required | As of early 2025, creating GPTs is limited to paid ChatGPT users and, in managed workspaces, to users with permission to create or edit GPTs. Free-tier users can use GPTs but not build them. |
When to add Actions (and when not to)
Actions are where most builders overreach. They let your GPT call external APIs – per OpenAI’s platform docs, they translate natural language into the JSON schema required for a REST API call. That’s powerful. It’s also a commitment: you’re now maintaining an OpenAPI schema, auth, and error handling.
The decision rule: add Actions only when the GPT genuinely needs live data or needs to do something outside ChatGPT – file a ticket, fetch an order status, hit your internal CMS. For anything that’s just answering questions about documents, knowledge files are enough. Actions are for things that change in real time; everything else is just extra surface area to break.
Pro tip: Before adding an Action, ask: “Could I solve this by pasting the data into a knowledge file once a week?” If yes, do that.
The harder question – and one the docs don’t fully answer – is where the real complexity boundary sits. At what point does a GPT with five knowledge files and explicit routing instructions actually need an Action versus just better-written instructions? That line moves depending on your use case, and finding it is most of the real work.
Iterate with evals, not vibes
The thing that separates a GPT you use daily from one you abandon is the testing loop. After every instruction change, run the same 10-15 test prompts and check the outputs against what you actually want. If the GPT gets worse, revert. If it gets better on 8 prompts and worse on 2, you’ve found a tradeoff to think about – not a clean win.
Most people skip this because it feels tedious. It is tedious. It’s also why some custom GPTs feel sharp and others feel like a generic chatbot wearing a costume.
FAQ
Do I need ChatGPT Plus to build a custom GPT?
Yes – building requires a paid plan (as of early 2025). Free users can chat with published GPTs but can’t create one.
Why does my GPT keep ignoring the PDF I uploaded?
Almost always because your instructions don’t tell it when to use the file. Try adding an explicit line like “Before answering any question about pricing, consult the uploaded pricing-guide.pdf and quote from it directly.” Uploading the file makes it available; the instructions make it actually get used. If it still ignores the file, try splitting the PDF into smaller, clearly-named text files – complex layouts can make content harder for the model to process.
Can I migrate a GPT to a different model later?
The model can be changed after the fact, but instructions tuned against one model’s behavior may need rewriting for another – outputs, tone, and how the model responds to formatting cues all shift. With GPT-4o retiring in April 2026, this is worth testing on your most-used GPTs sooner rather than later. Re-run your eval prompts on the new model and adjust from there.
Next step: Pick one repeatable task you did this week. Open chatgpt.com/gpts, go straight to Configure, and write the instructions for a GPT that handles that task. Test it with five real inputs before you do anything else.