I built a "Fear & Greed Index" for world conflicts with Claude Code: here's how the build actually went

https://preview.redd.it/gtjcxgpmqish1.png?width=857&format=png&auto=webp&s=6a48ed58ac0dfe9c627c6b7758ba0afa033dc846 I'm a geopolitics nerd and a web dev hobbyist, and I built Global War Index, a free 0–100 gauge (think Crypto Fear & Greed Index) that measures how tense the world looks based on international news coverage. It's a measurement of media coverage, not a prediction. What it does: every 4 hours it pulls headlines from 10 international outlets (BBC, Guardian, Al Jazeera, DW, France 24, Defense News…), classifies them with Claude, and scores them into a global index plus sub-indices (Nuclear, Hybrid, Great-Power, Civil War, Bio/Chem), regional hotspots, a continent map, a 90-day history, and the list of headlines behind each reading. There's also a free embeddable widget. Free to try, no signup: https://globalwarindex.com How Claude Code helped (the actual process) • Planning before any code. I started by asking Claude Code to write a development plan, not code. We went through 3 versions of it in a markdown doc: which indices to track, which sources had stable RSS/APIs, hosting options, update frequency, and a hard rule that no scoring logic ever reaches the browser (the public site only receives final numbers and headlines). Only when the plan was locked did we start coding, in numbered phases. • Phased build. Claude Code implemented it phase by phase, each one tested before moving on: Phase 0–1: Python project, SQLite schema, source ingestion (RSS + The Guardian's API) and clustering, so one story covered by 10 outlets counts as one story. Phase 2: headline classification with the Claude API (Haiku 4.5), with a rules-based fallback if the API fails. Phase 3: 0–100 scoring, plus a 90-day historical backfill using the Batch API to keep costs low. Phase 4–5: a static site, deployed automatically to Cloudflare Pages from a GitHub Actions pipeline that runs every 4 hours. 3. Debugging real-world mess. This is where Claude Code saved me the most time: Some feeds rejected our requests, so it diagnosed a User-Agent issue and fixed it. One source (Bulletin of the Atomic Scientists) returned 403 to any automated client, so it proposed and wired in a replacement (38 North). The historical data source kept timing out, so it added retries with backoff and a circuit breaker. It added post-deploy checks that verify the live site, the contact form and the embeddable widget actually work after every run. 4. Calibration. After the 90-day backfill, we used that history to calibrate the scoring, so the gauge moves in a sensible range instead of being pinned at 0 or 100. • Polish. A continent map (fun detail: we had to split Russia at the Urals so western Russia counts as Europe and Siberia as Asia), social preview images, share buttons, a strict Content Security Policy, and an embed builder page. My role: product owner. I decided what to measure, which categories and hotspots matter, and how it should look and feel, and I reviewed everything. Claude Code wrote the code and tests and handled most of the debugging. What I learned: the hard part wasn't the code, it was defining "tension" in a way that stays consistent. Having Claude Code write a plan first, and keeping that plan updated as a living doc, made a huge difference. Feedback very welcome, especially on sources (the mix leans Western) and on anything confusing in the UI. Happy to answer questions about the build.