model-cleanup

A linter for hardtech financial models. Point it at a founder's or portfolio company's model and it either cleans it up into something an institutional investor trusts on sight, or detects the modeling anti-patterns in it without changing the file.

Inspired by the no-ai-slop skill — same idea, applied to spreadsheets instead of prose: a fixed catalog of named anti-patterns, each with a fix.

• Fix (default). Audits the model, applies the mechanical fixes (font, color code, number formats), centralizes assumptions, adds sources, builds the missing analytics, and returns a cleaned copy plus a What changed section. The original is never overwritten.
• Detect. Names each anti-pattern that appears, points to the exact cell, quotes the value, and gives the fix in a few words — without touching the file.

Twenty named patterns across three groups:

• Formatting — hardcodes in black, formulas colored like inputs, mixed fonts, units typed into values, inconsistent number formats.
• Structure — the same assumption typed in five places, magic numbers buried in formulas, no single source of truth, unsourced inputs, no legend.
• Analytics — no unit economics, an asserted utilization with no breakeven, blended margins that hide the story, no capex intensity or payback.

The house color code: inputs bright blue #0000FF, same-sheet formulas black, cross-tab links bright green #00B050, external/warning red #FF0000. Body font Arial 10pt.

Works across hardtech domains — energy, compute/datacenter, manufacturing, robotics, materials, climate hardware.

• SKILL.md — the skill: two jobs, the pattern catalog, the workflow.
• eval.md — the checklist a cleaned model must pass.
• scripts/model_doctor.py — the audit/fix engine (audit and fix subcommands; requires openpyxl).
• references/ — the formatting, structure, and hardtech-analytics standards in full.

Download the repo and add it as a skill in Claude (Cowork → Skills → add), or drop the folder into your skills directory. Then just say things like "clean up this model" or "what's wrong with this model?" and attach the .xlsx.