STACK · AUG 11, 2026 · 5 MIN
How AI Rewrites Our Website From the Sales Calls That Closed
Three months into Clankers, AI has rewritten our website 11 times. Each rewrite pulls the language from that week's closed deals through the Fireflies and HubSpot MCPs, so the site says what actually made buyers say yes.
Three months into running Clankers, AI has rewritten our website 11 times. Every rewrite pulls its language from the sales calls that actually closed that week. The system has three parts: a context layer built from the Fireflies and HubSpot MCPs, a positioning reference file that Claude maintains from the closed call transcripts, and a weekly generation step that produces three landing page variants for me to pick from and deploy in Vercel in under five minutes.
Key takeaways:
- The context layer is call recordings plus deal data, connected to Claude through MCPs, filtered to deals that closed.
- Claude maps the positioning statements from each closed call into a markdown reference file it keeps as a source of truth.
- Each week, Claude generates three landing page variants from that reference and the current site. I pick one and deploy.
- The result is a website that says what buyers responded to, updated weekly, with minimal effort from a solo founder.
Most websites are written once, by the founder, from memory of what they think the pitch is. Then the pitch evolves on sales calls for months while the site stays frozen. This system closes that gap by making the closed calls the source of the copy. Here is how it works.
Why should sales calls write the website?
Because the calls are where the positioning gets tested. Every call is an experiment in which words land and which do not, and the deals that close are the experiments that worked. A website written from those transcripts is a website written from evidence.
The alternative, writing from instinct, produces copy that sounds right to the founder and is untested against anyone who actually pays. When the site and the winning pitch drift apart, the site is the one that is wrong.
How does the system work?
1. Build a context layer
We use the Fireflies MCP for call recordings and the HubSpot MCP for deal data. Fireflies gives Claude the transcripts; HubSpot tells it which deals closed.
Then we pull the closed deals weekly and read their call transcripts. The filter to closed deals is the important part. Reading every call would mix the winning language with the losing language and average them out. Reading only the closed ones biases the reference toward what worked.
2. Extract what language landed
I map the key positioning statements from each closed call into a markdown file that Claude stores as a reference. This file is the memory of the system. It accumulates the phrases, problems, and framings that appeared in conversations that ended in a deal.
Here is an example of three problems we position around on our website, all of which came out of this process:
- Lack of technical fluency on the team
- Existing data never leveraged with your LLM
- Bleeding efficiency through too much manual work
Claude found that when I consulted companies on these pain points, deals closed more often. I did not decide those were the three problems. The transcripts did.
3. Generate landing pages
I have Claude use the reference file and the current website to generate three landing page variants. The instruction is roughly:
Read the positioning reference and the current site. Generate three landing page variants that lead with the problems and language from the reference, keeping the structure of the current site where it works.
I pick one and deploy it in Vercel in under five minutes. The generation is automated; the choice is not. Three variants is enough to see real differences in emphasis without turning the pick into a chore.
What has it produced?
Eleven rewrites in three months, each one pulling the site closer to the language that closed the most recent deals. The practical effect has been more qualified leads entering the sales process, with minimal effort on my end as a solo founder. The site does the first round of qualification because it says the things buyers already told me they care about.
What does this look like for a brand or agency?
The mechanics transfer directly. Any team that records sales calls and tracks deals in a CRM has the two inputs. The context layer is the piece most teams are missing, and it is the same piece that blocks most AI use cases: the data exists, but it is not connected to the model in a form it can read. The marketing data foundation post covers what that foundation needs to look like before workflows like this one can run.
Once the context layer exists, the website is only the first thing it can rewrite. Sales decks, outreach templates, and onboarding emails all have the same problem of drifting away from the live pitch, and the same fix. The pattern of pulling structured data through MCPs into Claude and rendering a deliverable is the same one behind automated weekly reporting; only the sources and the output change.
Frequently asked questions
What is a context layer for sales copy?
It is the set of sources AI reads before writing: in this case, call recordings from Fireflies and deal data from HubSpot, connected through their MCPs. Without it, AI writes generic copy. With it, the copy is grounded in conversations that closed.
Why only use closed deals?
Closed deals are the ones where the positioning worked. Pulling language from every call would mix in the pitches that lost. Filtering to closed deals keeps the reference file biased toward what buyers actually responded to.
How often does the website get rewritten?
Weekly. The closed deals from the week are pulled, their transcripts read, and the positioning reference updated. In the first three months that produced 11 rewrites.
Does AI deploy the site automatically?
No. Claude generates three landing page variants from the reference and the current site. A human picks one and deploys it in Vercel. The choice stays with the founder; the drafting is what AI handles.
What does this require technically?
Fireflies and HubSpot connected to Claude through their MCPs, a markdown reference file for positioning statements, and a site that can be redeployed quickly. Vercel deploys in under five minutes.
Written by Sohun Sanka, founder of Clankers, an operator practice that builds AI systems and automation for social-commerce brands and agencies inside the tools they already use. This post expands on a LinkedIn post Sohun published on July 20, 2026.
Connect the calls to the copy
If your sales calls and your website have drifted apart, the context layer is the fix, and it is the same layer every other AI workflow in your business needs. The Clankers 90-Day Revamp builds it inside the tools you already run.