How I Built a Custom GPT for Webflow Client Onboarding That Saves Four Hours Per Project

Why Does Webflow Client Onboarding Eat So Many Hours?

Every time I start a new Webflow build, the first week looks the same. I send a welcome email, share a Notion intake doc, set up a shared Google Drive, ask the same fifteen questions, follow up because two answers are missing, write the project brief, and finally book the kickoff. By the time the design starts, I have already burned five or six hours that I cannot bill cleanly. Last month I worked with a Bengaluru founder who runs a small SaaS, and the intake alone took three rounds of email tag.

According to a 2025 Freelancers Union survey, freelance designers spend roughly thirty percent of project time on non-billable admin. For a fixed-scope Webflow build, that overhead quietly eats into my margin. I had been trying to shrink it for two years using templates and Notion AI. Then I built a Custom GPT for onboarding, and the time dropped from six hours to under two hours per project.

This post walks through what I built, what I fed it, where it broke, and how you can copy the approach for your own Webflow practice. I want you to be able to ship a working onboarding GPT this weekend.

What Is a Custom GPT and Why Did I Pick It Over Claude Projects?

A Custom GPT is a configured version of ChatGPT with a fixed instruction set, files, and tools that you publish inside OpenAI's GPT Store or keep private. As of OpenAI's late 2025 disclosure, the GPT Store had crossed three million Custom GPTs. I picked it over Claude Projects because most of my clients already pay for ChatGPT Plus.

I do use Claude Projects for long-form writing because Claude Opus 4.7 handles tone consistency better. But for onboarding, the daily workflow happens inside ChatGPT, which is where I already paste meeting notes and Loom transcripts. Switching tools for one workflow created friction. The GPT lives in the same sidebar I use for everything else.

The trade-off is that Custom GPTs have weaker memory than Claude Projects, so I had to be more deliberate about uploaded files. If your stack is built around Anthropic, Claude Projects with the same approach will work just as well.

What Did I Feed My Onboarding GPT to Make It Actually Useful?

I uploaded eight documents and wrote about 1,400 words of instructions. The documents covered my standard Webflow scope template, my discovery question bank, three sample client briefs from past projects with names redacted, my pricing tiers, my refund policy, and a CSS naming convention guide that matches how I structure Webflow class names. Each document is under fifty kilobytes so retrieval stays fast.

The instructions tell the GPT to ask the client one question at a time, never lists. It is told to confirm understanding by paraphrasing the last answer, then move on. It is also told to flag missing information in plain language rather than generating filler. The single biggest improvement came from one line: "Do not write the brief until the user types READY."

That stop signal prevents the GPT from drafting a generic brief halfway through the conversation. Founders type their answers, then say READY, and the GPT compiles a tight one-page brief I can hand to my designer.

How Does the GPT Actually Save Me Four Hours Per Project?

The four hours come from three places. The first is intake itself. Instead of writing custom emails and chasing answers, I send a single link to the GPT with my client's name in the system prompt context. The client talks to it directly, answers seventeen questions in twenty minutes, and exports a Markdown summary. That alone replaces about ninety minutes of back and forth.

The second saving is brief writing. The GPT produces a structured brief in my format: goals, audience, must-have pages, integrations, budget, timeline, and risk flags. I review, edit two sections, and ship it. That is roughly two hours of writing collapsed into thirty minutes of editing. The third saving is the kickoff call itself. Because the brief is already aligned with how I run projects, the call moves to decisions instead of discovery.

On three recent projects I tracked the time carefully and the savings were consistent: three hours and forty minutes on a Mumbai ecommerce build, four hours and ten minutes on a marketing site for a Dubai founder, and four hours flat on a portfolio relaunch for a Bengaluru consultant.

What About Privacy When Client Documents Go Into the GPT?

I do not upload any client documents to the GPT itself. The uploaded files are my own templates and redacted samples. Clients paste their own answers into the chat, and I keep training off using OpenAI's data controls toggle in the GPT settings. As of OpenAI's enterprise data policy update in early 2026, conversations in private Custom GPTs are not used for training when this toggle is set.

For sensitive clients in finance or healthcare, I do not use the GPT at all and run the onboarding manually. For everyone else, I include one line in my engagement letter that disclosures any AI tools I use during the project. The clients I work with appreciate the transparency more than they worry about the tool.

How Does This Compare to Notion AI or a Generic ChatGPT Thread?

Notion AI is great for writing and summarizing, but it cannot run a guided interview. It expects a finished document, not a turn-by-turn conversation. A generic ChatGPT thread can do the interview but forgets my instructions after a few sessions, so I had to re-prompt every time. The Custom GPT carries the instructions forever, which is the whole point.

I tested the same intake flow against Gemini 3 Pro's Gems and Perplexity Spaces. Gems worked well but my clients struggled with the Google sign-in flow if they used personal Gmail accounts that were not connected to Workspace. Perplexity Spaces was too sparse on multi-turn memory at the time I tested in April 2026. ChatGPT won on pure client friction.

How Do You Build This For Your Own Webflow Practice?

Open ChatGPT, click Explore GPTs, then Create. Give it a name like "Webflow Intake for Your Practice". Paste in instructions that cover three things: how it should ask questions, what brief format to produce, and how to handle missing data. Upload your own scope template, three sample briefs, and your question bank. Set conversation starters that mirror how a client would begin: "I want a new website", "I want to redesign", and "I want to add features".

Then test it as a client. Run through your own intake from a fresh browser. Note where the GPT drifts, where it pads, where it skips. Tighten the instructions. The second version is always sharper than the first. Mine is on version four and I still tweak it every few projects.

How Do You Know If Your Onboarding GPT Is Actually Working?

I track two numbers. The first is intake-to-kickoff time, which used to be eight days on average and is now two and a half days. The second is the number of clarifying questions I ask in the kickoff call. Before the GPT, I asked roughly twelve. Now I ask three or four because the brief is already complete. Both numbers come from a simple Airtable I log after every project.

If those numbers do not move in the first three projects, your GPT is too generic. The fix is almost always tighter instructions and better sample briefs. The data shows that according to Anthropic's 2025 agent evaluation guide, narrow scope and concrete examples beat clever prompting every time.

How Can You Set This Up This Week?

Block ninety minutes on Saturday morning. Spend the first thirty minutes collecting your three best past briefs and redacting client names. Spend the next thirty minutes writing instructions that describe how you actually interview clients, in your own voice. Spend the final thirty minutes building and testing inside ChatGPT. Send the link to one trusted past client and ask them to walk through it.

For the foundation of how I structure my intake before the GPT touched it, my walkthrough of a paid Webflow discovery call in Bengaluru covers the question logic I still use today. For a deeper comparison of when to pick Claude Projects versus Custom GPTs, my breakdown of Claude Projects versus ChatGPT Custom GPTs for Webflow walks through the decision in detail.

If you want help wiring this up for your own Webflow practice or want me to review your instructions before you ship them, reach out. I am happy to walk through it. Let's chat.