Case Study

Brand Pulse

The Form That Talks Back

  Web Design
  Content & Creative
    Systems & Automation

Proof Before Pitch

The Situation

  
Brand Pulse began with a question I kept circling: what can AI make that could not have existed before? Not faster versions of familiar work, but new experiences. If Colosseo was going to offer that kind of work, I wanted to build one for the studio first, where the risk was mine.

The answer started with an ordinary thing on the web: the onboarding form. A brand evaluation survey can ask a visitor anything, but it cannot respond. The visitor hands over what they know and gets a thank-you page. A read that reflected what they actually typed took a person and a discovery call. A language model changes that, for pennies per session, on a Squarespace site with no backend of its own.

I did not want another intake bot. I wanted an exchange, where a visitor gives a little and leaves with something useful: a read on where they stand and one thing to try this week, whether or not they ever hire the studio. There was no question count at the start, only an information problem: what did the tool need to know to say something worth reading, how little could it ask, and what was feasible to build? It also had two further jobs, capturing leads and showing what the studio can make. No client stood between me and the result, which also meant no one to tell me when I was wrong.

The model was the inexpensive part. The work was deciding everything around it.

From Six Questions to Three

Phase One - Listen

The first concept looked the way AI tools were expected to look: a standalone chat interface, a greeting screen, six questions. The greeting screen was the first thing cut, because the homepage copy already supplied the context and the screen was only friction.

Six questions became three, in an order drawn from the Listen phase of The M.O.: who you are, what is holding you back, what people should feel. The first is open and does the heavy lifting: “Tell me about what you’re working on — the real version, not the polished one.” “What you’re working on” holds for a ten-year company and an idea on a napkin alike. I rejected the more explicit “the more you share, the sharper your results” because it read like instructions. The other two offer set answers to choose from.

Those calls came from more than ten years of interface work and were made in-house, not through a research phase. Testing came later: my own runs with invented businesses, then a handful of people I showed it to.

Start here. You might be closer than you think.

Phase Two - Build

The First Line.

The input’s resting text carried a disproportionate load, because nothing downstream matters if no one engages. “Who are you and what do you do?” read like a networking name tag. “Tell me what you do. I’ll tell you what I see.” felt disingenuous, since the model does the seeing. “Your brand is already saying something. Want to know what?” assumed a brand that many visitors do not have yet. The line that held works as the next line of the headline rather than a competing hook. “Everyone has a story. Let’s find yours.” is followed by: “Start here. You might be closer than you think.” I tried adding “or further” and cut it. It added doubt about whether clicking was worth it, not useful tension.

The Reactions.

One model call at the end would have been simpler. It would also have been a form with a long loading screen. So a call follows each of the first two answers and returns one line proving the answer was read: “Starting from scratch isn’t a disadvantage. You get to build it right the first time.” The line streams in word by word, lingers for two seconds, and gives way to the next question. The dollar cost is negligible. The real price is latency and complexity, and I chose to pay it.

The Wait.

The read is held for about fourteen seconds at minimum, while a sequence quotes the visitor’s own answers back (“Still thinking about…”). This is deliberate theater. The intent is to signal that the result was considered, not dispensed.

The Read.

Four sections: Where You Stand, The Tension, The Opening, and One Thing to Try This Week. The brief to the model was specific: make them feel seen, name the tension they have felt but not articulated, point toward a direction, and do not solve it. The tool is meant to open a conversation, not replace one. One test run, for an invented bookshop-bakery, described the situation as “like you’re preparing for a dinner party no one knows they’re invited to.”

I considered a click-through reveal and rejected it. The power is in feeling seen all at once; piece by piece, it “feels like a slot machine, not a revelation.” So the sections stagger in together across three columns. Nothing is gated. Email is an optional copy, not the price of admission, and the card says plainly what made it: “Designed and guided by Colosseo Creative. Compiled by AI.”

Test run with an invented business.

The Placement.

I built Brand Pulse in working sessions with Claude. Claude wrote most of the JavaScript; the design, copy, and product decisions were mine. Those sessions included disagreement. Claude recommended placing the tool below the hero so the positioning could land first, a sound and conventional call. I overruled it. I wanted the tool to be the first thing a visitor saw and could do: an advertisement for the new kind of work this project set out to make. The risk was a chat window dominating the screen, so the resting state is a single line beneath the headline. The two act as one unit, and the headline yields only on interaction. Returning visitors who do not need the tool can skip it, and the hero still works for them.

The System.

The architecture stays small on purpose. Prototypes were React; what shipped is plain HTML, CSS, and JavaScript in one Squarespace code block, with no build step. A Cloudflare Worker hides the API key, rate-limits by IP, and accepts requests only from the Colosseo domain. A self-hosted n8n workflow logs each completed session and handles the optional email. Development testing cost less than a dollar in API calls. The model call sits behind the Worker rather than in the page, so moving to another model or provider would mean reworking that layer and retuning the prompt, not rebuilding the experience.

Every Run Reported Success

Phase Three - Sustain

When Anthropic retired the model Brand Pulse was calling, nothing crashed. The fallback I had written to keep the conversation moving kept it moving: anyone who used the tool got generic reactions and an empty results card. n8n reported every run as successful and logged rows of nulls.

I had seen this failure shape before launch. During testing I hit my own rate limit, and the tool quietly answered “Noted. Let’s keep going.” I treated that as a testing nuisance. It was a preview.

It took me longer than it should have to notice. Once I looked, a direct call to the Worker returned not_found_error for the model, and a one-line model swap restored the tool. Diagnosis took minutes.

A green automation run proves the pipeline moved data. It does not prove the product worked.

A fallback with no alarm hides failure from the one person who could fix it, and a failed run should never be logged as data.

What Changed

The Demo Did Its Job

Brand Pulse works, and it does what I built it to do: a visitor describes what they are working on and leaves with something useful, with no email required.

Its other job was showing what the studio can make, and that is where it has paid off so far. I have shown it to a few clients, and one has since come back to explore how something similar could work on their own site.

The conversation starts differently when the thing already works.

Lead capture is still ahead of it. Colosseo has grown organically rather than through advertising, so the homepage sees modest traffic, and most sessions so far have been mine or those of people I have shown it to. The next step is to measure starts as well as completions, so traffic and interest can be read separately.