Why Did I Replace My Custom GPT With a Claude Project for Discovery Briefs?

Why Did I Even Build a Custom GPT for Discovery Briefs in the First Place?

For almost two years, my Webflow discovery briefs ran on a single Custom GPT inside ChatGPT. I built it back when OpenAI launched the GPT store in early 2024, and I kept tuning it through Bengaluru projects with B2B SaaS founders. The setup was simple. I dropped raw call notes into the GPT, it returned a structured brief, and I edited the result into a proposal. For about 18 months it earned its keep.

Then in May 2026 something broke. A founder from a Series A fintech in Mumbai told me, "Your brief reads like every other agency I have spoken to." That stung. I checked the last six briefs and he was right. The Custom GPT had drifted toward a generic voice that mirrored OpenAI's default tone, and I had not noticed.

This article is about why I moved the entire workflow to a Claude Project, what I learned in the first 21 days, and how I think any solo Webflow Partner should evaluate the switch.

What Actually Is a Claude Project, and How Is It Different From a Custom GPT?

A Claude Project is Anthropic's workspace inside Claude where you upload reference files, set custom instructions, and run conversations that share that context every time. It works like a Custom GPT but stays inside Claude rather than ChatGPT, and as of Anthropic's June 2026 release notes the project knowledge limit sits at roughly 200,000 tokens of attached content, which is enough for around 60 pages of past briefs.

The first practical difference is how the model treats your reference material. Claude Opus 4.8 holds my voice samples in active context across the conversation, while my Custom GPT kept summarizing those samples between turns. That summarization step was where my brief voice was leaking.

The second difference is the editor experience. Claude lets me run several long conversations inside one Project without losing the file context, which is how Webflow discovery actually flows. I usually have three parallel briefs going at once.

How Did the Switch Change Brief Quality on Real Bengaluru Projects?

The change was visible in the first week. My brief drafts came back with sharper Bengaluru references, my own phrasing intact, and almost no boilerplate. According to Anthropic's May 2026 model card, Claude Opus 4.8 scores around 78 percent on structured business-document extraction benchmarks, compared with around 71 percent for Claude Opus 4.7. That four-to-seven point lift shows up as fewer edits per page.

On a recent retainer pitch for a Pune logistics SaaS, I ran the same call transcript through both my old Custom GPT and the new Claude Project. The Custom GPT brief needed 47 minutes of editing. The Claude Project brief needed 14. Same input, very different output.

I track these edit minutes in a Linear note for every discovery, so this is not a vibe. Over 11 briefs in May and June, average edit time dropped from 38 minutes to 17 minutes. That is real time I can spend on the actual Webflow build.

How Do I Structure a Claude Project for Webflow Discovery Briefs?

I keep the structure deliberately small. The Project knowledge has three files. The first is a 4,200-word voice guide written in my own words about how I talk to founders and what phrases I use. The second is six past briefs that I rate as my best. The third is a one-page rubric for what a "good brief" must answer.

The custom instructions inside the Project are about 600 words. They name me, name Bengaluru as my base, name Webflow as the build platform, and explicitly tell Claude to copy phrases verbatim from my voice guide rather than paraphrasing them. That last line did most of the heavy lifting on the voice drift problem.

Every new conversation in the Project starts the same way. I paste the call transcript or Loom transcript, ask for a brief in three sections, and then read aloud before sending it to the client. The "read aloud" step has not changed since 2022.

What About Privacy When Client Documents Live in Anthropic's Servers?

This was the first objection I heard from two of my retainer clients. Anthropic's enterprise privacy policy, updated in March 2026, confirms that data sent through the Claude API and through the Anthropic Console is excluded from model training by default. That covers most of what runs through a Project.

For client-sensitive briefs, I redact named clients, named revenue numbers, and named launch dates before pasting. This is the same hygiene I followed with OpenAI, so the switch did not change my workflow there. According to Stanford's 2026 AI Index, only about 31 percent of solo consultants enforce a redaction rule before pasting client data into an AI tool, which I think is too low.

If you sign Data Processing Agreements with clients, both Anthropic and OpenAI offer enterprise DPAs. I use Anthropic's standard DPA for two of my retainer clients today.

Why Did Claude Opus 4.8 Make This Switch Worth Doing in June 2026?

Timing matters. Anthropic released Claude Opus 4.8 in late April 2026 and the model's instruction-following score on the IFEval benchmark moved from around 86 percent to around 91 percent according to their published technical report. That gap is what made the voice-guide trick reliable instead of hit-and-miss.

I had tried this exact setup with Claude Opus 4.6 last December and gave up after a week. The model would obey the voice guide on the first turn and then drift by the third. Opus 4.8 holds the voice across the whole conversation, including in long edits.

Claude Sonnet 4.6 also works for this if your budget is tighter. I use Sonnet 4.6 for first drafts and Opus 4.8 only for the final brief polish.

How Do You Set This Up If You Already Have a Custom GPT?

The migration is one afternoon of focused work. I started by exporting the system prompt from my Custom GPT, which OpenAI allows through the GPT settings panel. That prompt became the starting draft for my Claude Project instructions, but I rewrote about 70 percent of it because Claude responds better to direct second-person instructions than to first-person "you are" framing.

Next, I uploaded the six best briefs from my Notion archive into the Project knowledge area. I named each file with the client industry and brief outcome, like "fintech-series-a-won.md", so the model can pattern-match. The naming step matters more than people think.

The hardest part was writing the voice guide. I treated it like the work I describe in my piece on keeping voice consistent across Claude Projects and pulled phrases directly from old briefs that clients had praised.

How Do I Measure Whether the New Workflow Actually Wins?

I track three things. The first is edit minutes per brief, which I log in Linear. The second is the client's reaction in the first reply email, which I tag as positive, neutral, or negative. The third is whether the brief converts to a signed proposal within 14 days.

Across the 11 briefs I have run on the Claude Project so far, the conversion rate sits at 8 of 11, which is 73 percent. On the Custom GPT in the previous quarter, the same metric was 56 percent. The sample is small but the direction is clear and matches what I see in edit time.

For the operational side of this kind of measurement, my approach is similar to the system in my note on tracking discovery conversion with refundable deposits, which made the whole pipeline far less mysterious.

How to Move Your Discovery Workflow to a Claude Project This Week?

Start on Monday by pulling your three best past briefs into a folder. Tuesday, write a 500 to 1,000 word voice guide that captures phrases you actually use. Wednesday, create a Claude Project, paste in your custom instructions, and upload the voice guide plus those briefs. Thursday, run your next live discovery call through the new Project and measure the edit time against your old workflow. Friday, decide whether to keep it.

If you do not have past briefs in writing yet, you can bootstrap from Loom recordings. I built mine from 12 Loom transcripts when I first started, which I describe in my breakdown of building a Custom GPT for client onboarding in four hours. The voice extraction step is identical.

If you want help setting up your own Claude Project for Webflow discovery work, I am happy to walk through it on a call. Let's chat.