← ClaudeAtlas

deal-evolutionlisted

The governing way Underwriter OS produces ANY analysis (feasibility, deal, market study, memo, model, site/market read). From ONE request, automatically evolve the work through V1 → V2 → V3 — drafting, then hardening, then finalizing — without the user asking again. Always invoked for analytical deliverables. The user asks once; you deliver the FINAL.
Fusion-Data-Company/underwriter-os · ★ 1 · AI & Automation · score 74
Install: claude install-skill Fusion-Data-Company/underwriter-os
# Deal Evolution — V1 → V2 → V3 (automatic, one prompt) **The skill is the evolution.** When someone asks for an analysis, you do NOT hand back a first draft and wait. You run the whole arc yourself — draft it, attack it, finish it — and deliver the V3. The user prompts once. ## The discipline (this is the part that matters) - **Finish each version before you plan the next.** Do not pre-plan all three. Produce V1 completely → THEN look at what you actually made → THEN decide how to harden it → produce V2 → THEN critique again → produce V3. Plan the next phase only from the finished prior one. (Don't over-plan up front — build, then critique, then build.) - **Each version must already carry full rigor** (see `underwriter-os/reference/rigor-standard.md`). V1 is not "sloppy" — it's the honest first pass with best-available data. V2 and V3 raise verification and depth, not basic competence. - Keep going until V3 meets the bar. Don't stop at V1 or V2 and ask "want me to keep going?" — you already know they do. ## V1 — DRAFT (get a real, defensible answer on the page) - Build the full structure and a first-pass answer using the best data you can pull now (parcel data if connected, FRED/Census if connected, web/Tavily/Exa, `config/firm-defaults.md` for the firm's calibration). - Compute the core metrics. Label every estimate as an estimate. - **End V1 when:** the deliverable is complete end-to-end and internally consistent — a real answer, not an outline. ## Self-critique gate (