The Discovery Call Notes Folder I Stopped Opening on Monday Mornings
For most of 2025, my Monday mornings started with a folder of half-typed discovery call notes and a half-finished Figma sitemap. I knew the founder's product, I had the recording, I had the transcript, and I still spent two hours staring at a blank Webflow page tree wondering where to start. In April 2026 I changed one thing in my workflow. I started feeding the discovery transcript to Claude Sonnet 4.6 with a sitemap prompt I had refined over twenty calls. Now the Monday-morning blank page is gone, and the first draft of the sitemap sits in my Webflow site map plugin within fifteen minutes of the call.
This is not "AI does my job" hype. It is a small, very specific use of Claude Sonnet 4.6 that has compounded across thirty-one Webflow projects since April. According to Anthropic's May 2026 update note, Sonnet 4.6 improved on long-context reasoning over Sonnet 4.5 by a measurable margin on the SWE-bench Verified set, and the same gains show up when you ask it to structure a sixty-minute conversation into a navigation tree. I want to walk through exactly how I use it.
The article covers the prompt structure, what I keep out of the model's context, how the first draft compares to what I ship, and what I tell clients about the role of Claude in my process.
What Is a First-Draft Sitemap and Why Does It Matter in 2026?
A first-draft sitemap is a navigation tree, page-by-page, with a one-line purpose for each page, generated before any visual design work begins. It matters in 2026 because Webflow projects now compete with AI-built sites that can ship a passable site in a weekend. The differentiator is not speed, it is the quality of the page architecture and the clarity of each page's job.
A bad sitemap turns into expensive redesigns inside Webflow. A good first-draft sitemap turns into a confident scope, a tighter quote, and a build that does not balloon. Webflow's June 2026 internal study, shared on the Partner forum, found that Webflow Partners who lock sitemap architecture before design start ship in roughly thirty percent less calendar time on average. That maps to my own data across the last forty projects.
I used to build sitemaps by intuition. Now I anchor the first pass to what the founder actually said on the discovery call. Claude Sonnet 4.6 is the bridge between the transcript and the tree.
How Do I Set Up the Claude Project Before the Discovery Call?
I run discovery calls inside a Claude project named "Webflow Discovery." The project has three persistent files. The first is my own sitemap framework, a one-page document that lists the canonical page types I build into B2B SaaS Webflow sites, with the criteria for including each. The second is a redacted example transcript with the corresponding sitemap I shipped, so Claude can pattern-match. The third is a style guide for how I write page purpose statements.
I add the founder's website URL and one-paragraph brief at the start of the project. I do not paste their pitch deck. I have found that decks bias Claude toward marketing-heavy page structures that do not survive the build. The transcript and the website are enough.
The Claude project also has a memory note that lists three pages I almost always end up building but that founders rarely mention: a comparison page, an integrations page, and a security page. Anthropic's memory tool in Claude, generally available since May 2026, makes that persistence reliable across sessions.
What Prompt Do I Use to Turn the Transcript Into a Sitemap?
The prompt is three paragraphs. The first paragraph names the audience, the founder's product, and the goal of the website. The second pastes the discovery call transcript. The third gives Claude its task in my voice.
The task paragraph reads roughly: "Read the transcript with the sitemap framework attached to the project. Propose a complete first-draft sitemap as a nested list of pages with a one-line purpose for each. Flag any pages where the founder said something on the call that contradicts a default I usually include. Keep the structure to a maximum of fifteen primary pages. Do not invent product features that were not in the transcript."
The "do not invent" line matters. Without it, Sonnet 4.6 will helpfully add a "Customers" page even when the founder said on the call that they cannot use logos publicly yet. The model is helpful by default, and you have to constrain helpfulness on a discovery call output.
How Good Is the First Draft Compared to What I Ship?
The first draft is roughly seventy percent of the final sitemap. I keep about seventy percent of the page list, rename ten to fifteen percent for voice, and add or remove the rest based on my own judgment after a second pass.
What Claude is consistently good at is catching pages I would have forgotten in the first hour. On a recent fintech project, the founder mentioned compliance certifications three times on the call. I would have batched that into the footer in my first pass. Claude proposed a Trust Center as its own page, with a one-line purpose that read "establish regulatory credibility for procurement reviewers." That was the right call, and it became the highest-converting page on the site after launch.
What Claude is consistently bad at is hierarchy. It will give you a flat list of fourteen pages with no nesting and call it a navigation. I always restructure the tree into primary and secondary groups myself before I take anything to Figma or Webflow.
Should I Send the Claude Output to the Client Directly?
No. I never send the Claude output to a client. The first draft is a working artifact, and showing a client a raw LLM response on a sitemap erodes trust for two reasons. First, it gives the impression that they are paying for AI, not for my judgment. Second, the output language is too generic, even from Sonnet 4.6. Phrases like "establish credibility" and "drive qualified leads" need to be rewritten in the client's language before they go on a sitemap document.
I rewrite the sitemap inside FigJam, paste it into a Loom video where I walk through each page in my own words, and send the Loom with a tight summary. The client sees my reasoning, not Claude's reasoning, and that is what they are paying for.
How Do You Handle Confidentiality and the Transcript in 2026?
Confidentiality is the question every founder asks once they realize I run transcripts through Claude. I tell them three things. First, I use Claude with the Anthropic Zero Data Retention setting enabled, which means Anthropic does not retain the prompt or completion. Second, the project workspace is scoped to me alone, with no team access. Third, I redact any specific competitor names, dollar figures, and internal customer names from the transcript before it goes into the project, using a Find-and-Replace pass in TextEdit.
The redaction takes me four minutes. It is a small price for the trust it buys on the call when I explain my workflow. If a founder is uncomfortable, I run a manual pass without Claude. So far in 2026, only one founder out of forty-one has asked me to skip the AI step.
What Models Did I Test Before Settling on Claude Sonnet 4.6?
I tested GPT-5.4 from OpenAI, Gemini 3 Pro from Google, and Claude Sonnet 4.6 from Anthropic on the same three discovery transcripts in April 2026. GPT-5.4 produced the most polished prose for page purposes but added the most invented pages. Gemini 3 Pro was the fastest but flattened the hierarchy worst. Claude Sonnet 4.6 hit the best balance of constraint compliance and structural usefulness.
I have not retested Gemini 3 Ultra since its general release in late May 2026. It is on my list. For now, Sonnet 4.6 stays.
How Do You Hand the Sitemap to Webflow Without Losing Context?
I hand the sitemap to Webflow through three steps. First, I create page placeholders in the Webflow Designer with the page name and a single H1 placeholder that uses the purpose statement. Second, I add Webflow page settings with a SEO description that is a polished version of the same purpose. Third, I log the rationale for any page deviations into the project Notion so future-me knows why a page is there.
If you want a closer look at how I keep AI-generated context attached to Webflow client work, my piece on why I replaced my custom GPT with a Claude project for discovery briefs walks through the project structure, and my note on how I use Anthropic's memory tool in Claude for Webflow client context covers the persistence layer that makes this scale across thirty clients.
How to Try This on Your Next Webflow Discovery Call
Set up a Claude project before your next discovery call. Add your own sitemap framework and one good past example to it. Run the call, get a transcript from Otter or Granola, paste it into the project with a three-paragraph prompt, and ask for a first-draft sitemap. Spend twenty minutes editing the output instead of two hours staring at a blank page. Compare your final sitemap to the first draft and write down where Claude helped and where it hurt.
You will not need this on every project. For a one-page Webflow landing page, the overhead is not worth it. For anything with more than five pages, it has saved me time on every single project this year.
If you want help setting up a Claude project that works with your Webflow practice, reach out and I will walk through mine with you. Let's chat.
Get found, cited and the back office automated
Let's make your site the source AI engines quote and wire up the systems behind it.
Read more blogs
Let's get your website found and cited by AI
Tell me what you're working on, whether AI search is skipping your product, your back office is buried in manual work, or you need a build that does both.