Scout

Company insights, right in Slack.

Scout was a teammate in Slack I designed and built for Klue’s sellers.1 It turned a sales call into insights delivered in the chat sellers already had open.

Project facts

Where
Klue
When
Q2 20251
Role
Led the design sprint, then designed and built Scout
Built with
Replit, JavaScript, GPT‑4.1, Slack canvases, Figma2
Team
Sprint: the design team and the VPs of Product and Engineering. Scout: me.3
Shown to
Board of directors
About Klue
Software that helps companies gather competitive intelligence and share it with their sales teams
Scope
The sprint, the Figma prototype, the content and canvases, the prompts, the bot, and the live demo for the board

15

scripted steps the demo could follow, with open-ended questions woven in.

10

prompts in the pipeline, one for each step from the call to the insight.

1

sales call was used to generate every insight Scout delivered.

The problem

Sellers wouldn’t open another app.

Klue’s insights lived in its own app. To use them, people had to leave the conversation they were in and go looking. Before betting on an overhaul of the product, Klue had one question to answer: could it deliver the same value inside another tool’s interface?

A Klue insight for Render Services Inc. with its summary, "Pentera is back from the dead.", highlighted in yellow, next to the Ask Klue panel. In front of it is a Slack sidebar for the Rapid7 workspace. A red line runs from the #sales channel toward the Klue insight, and an × partway along marks that the two aren't connected.

Leaving Slack to find an insight. Klue had the answer; sellers had to open Klue to see it.

What I owned

I led the sprint, designed it, built it, and took it to the board.

I carried Scout from a sprint idea to a working bot in front of Klue’s board.

The solution

Every message, an insight.

An insight was a statement, the intel behind it and the action to take. Scout delivered them in Slack, with the documents behind each one a click away. A mockup couldn’t show what it was like to ask it something, so I built it, and it answered back.

A Scout demo brief for a Thursday 10am demo with Elena Wu from Render Services Inc. It names two competitors, Wiz and Pentera, links talking points, and lists three files. Around it are the pipeline documents it was built from: the SDR call transcript, competitor and selling-to prompts, SDR and AE qualifications, the AE synopsis and talking points. Thin blue lines run from each document to the part of the brief it fed.

Every line of the briefing, traced back to a prompt. The meeting, the buyer, the competitors and each canvas came from prompts that read one sales call.

Underneath, part one: the knowledge

Knowledge derived from one call.

Every insight Scout delivered came from one sales call. Specialized prompts read only the transcript and pulled out the buyer’s needs, the competitors, who it was selling to and the pain points, as structured data. Qualification notes, the value wedge and the talking points were built on top, and every insight drew on all of it.

The layers under one message. The call: a sales call transcript. Extracted: buyer needs, competitors, the account, pain points with a confidence and a status. Written for the deal: SDR qualification, AE qualification, talking points. What the seller saw: the message in the thread, and its canvases.

The layers under the briefing message, from the call transcript up.

Underneath, part two: the proof

Feasibility, proven one prompt at a time.

Building the pipeline was conversation design and feasibility at the same time. Each step was a real prompt that generated real answers based on real content. I used that content for the Figma demo, and showed engineers how I structured the pipeline to generate all the data that Scout needed to give good answers.

The pipeline in build order. The call transcript feeds four steps: buyer needs, pain points, selling to and competitors. Together they make the discovery notes, which lead to AE qualification, value wedge, opportunity synopsis and talking points. Those four make the insight, which becomes the Figma demo’s content, then Scout’s data, then Scout in Slack, shown with Casey’s avatar.

Each step proven before the next. Outputs that held up against the call became the Figma demo’s content, and then Scout’s data.

Tough calls

Only the conversation was real.

I had to move quickly on this one. So a majority of the concept stayed in Figma. And a vision video had to wait. I built what mocks couldn’t show: talking to it.

  • The bot first, the video after.

    A video couldn’t take a question from the board, so I built Scout first, then filmed it to show everyone else.

  • Guided, not open-ended.

    The board demo had to land the same way every time, so trigger words moved Scout through a set story. Questions off that path still went to the model.

  • Slack only.

    The board saw Salesforce, Gong and the web as designs; the conversation would have worked the same way in each.

  • The call as the only source.

    Each prompt read the call transcript and nothing else. That limited what it could say, but every claim could be checked against the call.

How Scout handled a message. 1, a message lands in Scout’s thread, and Slack passes it to the bot, which ran on Replit. 2, is it a scripted step? Scout checks it against its scripted commands, each with its trigger phrases, its prompt file and whether it attaches a file. If yes, 3 to 5: the step’s prompt loads and the reply is built from it, canvases included. If no: 3, classified (a question about the deal, a follow-up, how it knows something, help, a person, or something else); 4, it gathers context from the thread and the sandbox data; 5, a model writes the reply in Scout’s voice. 6, back in the thread, with its canvases.

How Scout handled a message, from the thread and back.

The outcome

Klue started building for chat.

In May I ran Scout live for Klue’s board of directors. It landed at a time when there was scrutiny over the direction the company was taking. It helped kick off Klue’s plan to deliver insights inside the tools its customers already use.

After the board demo, a timeline. May: Scout in Slack, a design-led prototype, live for Klue’s Board. June: Compete Agent, the launch of a clickable feature that generates insights. Early July: Chat Assistant, a shift in marketing to an assistant concept. August: a deal tip in a Slack DM becomes a requirement, then active development. A bracket from May to August reads: same idea, three months later.

Klue launched Compete Agent in June as a point-and-click tool and, by early July, marketed it as a chat assistant. By August, one of the CEO’s required flows was Scout’s core idea: a seller getting a deal tip in a Slack DM. Chat development continued from there.

15

scripted steps the demo could follow, with open-ended questions woven in.

10

prompts in the pipeline, one for each step from the call to the insight.

1

sales call was used to generate every insight Scout delivered.

Klue’s Compete Agent page on klue.com: Your competitive intel operative. A chat window answers a seller’s question about a competitor, with follow-up questions, between two illustrated agents. Below: Research Analyst and Competitive Deal Assistant.

Klue’s Compete Agent, on klue.com. Sellers ask it about the competition and get answers back.

My take

Sometimes you have to build the argument.

I’d long believed insights should arrive as documents, not screens. Scout showed it, live, the board agreed and the company shifted direction.

Questions? Answers.

  • What was Scout, exactly?

    A working prototype of a teammate in Slack for sellers. It turned a sales call into insights and the documents behind them, and it ran live for Klue’s board on a scoped set of sandbox data.

  • Did you design it, or build it?

    Both. I led the February design sprint with Klue’s design team. Then I designed and built Scout myself: the prompts, the content they generated, the scripted story and the bot. In May I ran the demo.

  • How did you know it could be built?

    Building it was the test. Each step of the pipeline was a prompt run on the sales call. Where its output held up, that piece was proven, and the next was built on top. When the pipeline had written all of the demo’s content, the design had been shown to work, and the first step had already worked on real discovery calls.

  • How long did it take?

    The sprint and a first Figma demo were in February 2025. The pipeline came next, and its content went into a refined Figma demo. I built the bot in May and tested it over two weeks, and ran it for the board the same month.

  • How is Scout different from futrbot?

    Scout was a Klue prototype, for Klue’s board, in 2025. It brought competitive intelligence to sellers in Slack. futrbot came a year later: a personal project built on OpenClaw.

  • What happened after the demo?

    Klue’s engineers began building chat-based infrastructure. Scout was the prototype; the product work that followed was theirs. The discovery-notes step kept running on real calls: a Zap posted notes to Slack after each new discovery call.

Next project

Search design.

Everything on a competitor, from a single search.

  1. February 2025: the design sprint. May 2025: the bot, built and run live for Klue’s board of directors.
  2. A Slack bot built for the demo, on OpenAI’s GPT-4.1, running on a scoped set of sandbox data.
  3. Scout: designed and built by me. The sprint: Klue’s design team, with the VPs of Product and Engineering and a temporary design manager.

The deal, the people and the data in Scout’s demo were simulated, for the demo.