War Room

Inspiration

Freight brokers win business by understanding a shipper before the first call: what it ships, where, how often, and who moves it today. That research is slow, varies from rep to rep, and usually lives in one person's notes. Our own timed attempts at a full prospect brief by hand averaged 55 minutes. The Neeley AI Forward case challenged us to turn scattered public signals into a brief a rep could read in two minutes and actually trust. We wanted to build something a sales team could reuse every day, not a one-off chat.

What it does

War Room turns a company name into a one-page, source-cited freight sales brief with a Sharpest Signal, call openers, and three things to check before dialing. Every line is tagged verified (tied to a dated public source) or inferred (with a reason and a confidence level). The brief computes three scores, Opportunity, Confidence and Freight Complexity, which set call order and how far to trust the brief. An interactive site shows the research run, the receipts behind every claim, a team table with PDF export, and gauges that count down to when each brief goes stale. If a company isn't in the library, the site gives you one prompt that carries the whole skill, so anyone can run it in Claude, even on a phone, and paste the result back in.

How we built it

The core is a reusable Claude skill: a plain-language instruction file that sets the research rules, the verified and inferred tagging, the scoring inputs and a shared JSON format. A single scoring module computes every score from that JSON, so nothing is typed in by hand, and automated tests check it against independent recomputation. The site is a static Next.js app with React Three Fiber scenes, built so every number on screen traces back to a brief or to our timed runs. The 3D scenes use public-domain and CC0 assets, including real USGS and SRTM elevation and NASA Blue Marble imagery for the map. Everything runs without a backend, logins or secrets.

Challenges we ran into

The AI was confidently wrong more than once: it dated a Dell signal by the quarter-end instead of the release date, and an early BlueLinx draft named a state from a guess. An early search also returned ten-year-old filings that would have misstated the company's network. Each mistake became a rule, such as dating signals by publication and keeping inferred geography regional, plus an automatic check that runs on every brief. Keeping the scores honest was hard, because they are heuristics, so we show the math and name the one judgment call that would change each score. Making rich 3D scenes work on phones for people scanning our QR code took its own round of fixes.

Accomplishments that we're proud of

A full brief now takes about 10 minutes, a 5-minute skill run plus a 4-minute human check, compared with 55 minutes by hand. Every claim in every brief can be clicked back to its source, and inferences are never presented as facts. We documented three real AI errors and showed how code catches pattern errors while people catch judgment errors. The library holds 18 companies, and anyone can add more without installing anything. We are proud the site says plainly what the tool can't do: see real shipments, or measure freight with its scores.

What we learned

AI is fast at gathering public signals but needs guardrails, because it can sound certain while being wrong. Requiring a dated source for every verified line and a reason for every inference made the output far more trustworthy. A short, focused human check of three claims beat trying to review everything. Designing for the team, with a shared format, refresh dates and corrections that travel with the brief, mattered as much as the brief itself. Being honest about limits made the tool more credible.

What's next for War Room

Next we want to connect briefs to a CRM so they live where reps already work. We would run the skill on a server so a brief can be generated with one click instead of copy and paste. We want to validate the score weights against real call outcomes rather than judgment. Automatic stale-brief alerts from new filings, facility announcements or logistics job postings would keep research current. Finally, we would test it with working freight reps to measure the time saved and the conversations it earns.

Built With

• claudecode
• claudeskills