Weather text generation

A job advertised as 50% science, 50% programming — in practice it was close to 100% programming, and it's what completed my move from scientist to programmer. The main work was turning structured weather data into natural-language forecast text under hard constraints, for example a forecast that had to fit inside a single SMS message. This was pre-LLM, rule-based text generation: fixed templates and branching rules, not a statistical model. The diagram below runs one forecast through that pipeline — watch the draft text blow through its 160-character budget and get trimmed back down to fit.

1. Data in — Keswick, 06:00 forecast run
2. Rain prob 80% → include shower clause
3. Gust 40mph → include wind detail
4. Draft assembled — 186 chars
5. Over budget: 186/160 — trimming required
6. Drop gust clause (lowest priority) → 162 chars
7. Still over: 162/160 — simplify further
8. Wind clause shortened → 151 chars
9. Fits: 151/160 — approved for send
10. “Keswick: cloudy with heavy showers this morning, 80% chance of rain, easing during the afternoon before a clear evening. Max 14C min 6C. Wind SW 25mph.” — SENT