灰灰

I built more 3-statement models by hand at Morgan Stanley than I'd like to admit. Claude in Excel is a game changer .... here's how I build any model in 5 passes: The process is all about context window management to minimize hardcodes + hallucinations and maximize control over model style. (1) Gather the sources. Filings, earnings transcripts, raw financials and any model or template I already have. I have it make a clean folder I point Claude to. (2) Turn the sources into a JSON or .md input. A short inputs file with historicals, segment KPIs, accounting definitions and operating drivers. This step converts many megabites of PDF's into a liteweight doc for RAG. (3) Build the structure (in your firm style). Building the shell lets you reverse engineer your firm model style .... which Claude will never have context for. (4) Add the forecast. I take the same historicals JSON or .md and run a deep research workflow to fill in those parameters but for forward assumptions. Obviously with sources so I can check assumptions. Then I layer in that file to project out forward assumptions. (5) Audit it in a fresh chat. New chat is necessary to hard reset the context window (improves quality). A simple prompt even to check for hard + soft errors can rank issues by severity (eg broken formulas and things that don't tie to assumptions that look off) If you're really advanced at Claude Code - you can wrap this into a programmatic skill package in your firm build style .... maybe I write a thread about that next time. In the meantime, I've put together this AI for Excel build guide here: davewang.ai/resources/how-to…