How to Create a Custom GPT for Your Agency (Step-by-Step)

Agencies are one of the clearest use cases for custom GPTs — your knowledge already lives in documents: brand guides, SOPs, past proposals, client briefs. The problem isn't figuring out how to create a custom GPT for your agency. It's that most teams spend hours wrestling with file prep and still end up with a GPT that confidently retrieves the wrong thing.

Here's how to do it right, from scoping the job to testing it with real questions.

What an Agency Custom GPT Actually Does

Before touching the build tab, lock down the job description. Custom GPTs in agencies typically serve one of four roles:

1. New client onboarding assistant — absorbs the brief, brand guide, and competitive landscape. Answers internal questions about the account so the team can ramp up faster without pulling a senior person into every briefing conversation.
2. Internal SOP assistant — answers workflow questions from your documented processes. Cuts the time senior staff spend re-explaining how things work to every new hire.
3. Proposal generator — trained on past proposals, case studies, and service-line descriptions. Drafts first passes on scope-of-work documents so your account team isn't starting from a blank page.
4. Client deliverable GPT — a custom GPT you hand to a client, pre-loaded with their brand standards and reference documents. They use it; you maintain it.

Pick one per GPT. A single GPT that covers all four does none of them reliably. Focused retrieval scope means the model has less to sort through — and the answers are sharper for it.

What to Put in the Knowledge Base

ChatGPT custom GPTs accept up to 20 files and 512MB of combined storage. That sounds like a lot until you realize a single uncompressed brand guide with embedded images burns through the cap quickly. Here's what actually belongs in each type of knowledge base:

• Onboarding GPTs: client brief, brand voice guide, audience personas, competitive summary, and the campaign brief template your team uses internally. Skip financial docs, contracts, and anything with confidential pricing — the model doesn't need it and you don't want it exposed.
• SOP assistants: your actual process documentation — workflow guides, naming conventions, QA checklists, brief templates. Granularity matters. "Follow company standards" in a process doc is useless context. "Submit to Google Drive → Client Name → Deliverables → [Month-Year]" is not.
• Proposal GPTs: 5–10 past proposals from different engagement types, a service-line description document, and a case study library in plain text. Avoid slide decks — ChatGPT retrieves text, not slide layouts.
• Client deliverable GPTs: the client's brand guide, content style guide, and any reference files that directly answer the questions their team will ask. Keep it narrowly scoped to their specific context.

Leave out anything with complex multi-column layouts, image-heavy PDFs, or files that are mostly charts with no surrounding text. The retrieval layer reads text — it doesn't interpret images of tables.

How to Prep Your Files Before Uploading

This is where most agency GPT builds fail. A PDF that looks clean in Acrobat is often unreliable for retrieval. Headers and footers repeat on every page and inject noise. Page numbers get included as standalone fragments. Multi-column layouts extract left-to-right across columns instead of column-by-column, breaking the logical flow of the content.

Before uploading any file, run a cleaning pass:

1. Extract the text from each PDF. Knowledge Builder Pro does this automatically — upload the file, download a clean TXT or Markdown version with headers, footers, and page numbers stripped. Your files are processed in-memory and never stored.
2. Check the first and last page of the extracted output. If you see repeated watermark text, garbled column order, or page numbers mixed into the content, clean those out before uploading.
3. Name your files descriptively. ChatGPT uses file names as part of retrieval context. brand-guide-acme-corp-2026.txt retrieves better than final_v3_USE_THIS_ONE.pdf. Rename every file to describe its contents plainly.
4. Split large documents by section. If a single document is over 50 pages, break it into logical sections before uploading. Retrieval from a 200-page PDF routinely misses specific sections because the relevant chunk doesn't score high enough in the retrieval pass.
5. Default to plain text formats. TXT and Markdown retrieve more reliably than PDF for text-heavy documents. Reserve PDF for files where visual formatting matters — like a visual brand guide where layout is part of the content.

Step-by-Step: Building the Custom GPT

Once your files are clean and named, the build takes under 30 minutes:

1. Go to chatgpt.com → Explore GPTs → Create.
2. In the Configure tab, write a system prompt that describes the GPT's job in one clear sentence. Example: "You are an onboarding assistant for the Acme Corp account. Answer questions about the client brief, brand voice, and campaign strategy using the uploaded reference files."
3. Upload your cleaned files in the Knowledge section. Stay within 20 files and 512MB.
4. Under Capabilities, turn off Code Interpreter unless your use case specifically requires it. For most agency knowledge GPTs, leaving it on increases hallucinations without adding value.
5. Test with 5–10 specific questions the actual users will type — not "summarize the brand guide." Ask exactly what a new account manager would ask: "What's the tone of voice for Acme's B2B audience?" or "What's our naming convention for client deliverables?"
6. If the GPT returns generic answers or invents specifics that aren't in the files, the problem is almost always file quality — not the prompt. Iterate on the files before you touch the system prompt.

Common Mistakes to Avoid

Building one GPT for every job. A custom GPT with 20 files spanning proposals, SOPs, client briefs, and competitor research retrieves poorly from all of them. Each use case deserves its own focused knowledge base. If your agency has five distinct workflows, that's five GPTs — not one.

Uploading slide decks. PowerPoint and Keynote exports convert to PDFs that are mostly images. The text on a slide is 10–20 words surrounded by layout data. Retrieval from slide decks is unreliable by default — convert the key content to a structured document before uploading.

Skipping the naming pass. Generic file names (Document1.pdf, final_v2.pdf) reduce retrieval accuracy. The file name is part of the retrieval signal. Take three minutes to rename every file before upload.

Testing with vague prompts. "Tell me about the brand" is not a useful test. Test with the exact queries your team or clients will actually type. If the GPT fails those, it fails in production — no matter how good the system prompt looks.

Treating the 20-file ceiling as flexible. It isn't. If your knowledge base outgrows 20 files, consolidate related documents into single well-structured files, or split the GPT into two focused tools with distinct knowledge sets. Trying to pack everything into one GPT beyond that limit means some files get dropped silently.

Wrapping Up

Knowing how to create a custom GPT for your agency is only half the work. The other half is building a knowledge base that retrieval can actually use. Most agency builds stall on file prep — converting PDFs to clean text, stripping layout noise, naming files so ChatGPT can surface the right context.

Knowledge Builder Pro handles the prep automatically. Upload your agency documents — brand guides, SOPs, proposals, client briefs — and get clean, structured output ready to upload to any custom GPT or Claude Project in seconds. No files stored after processing.