Theory Sweep — AI skill for Claude Code

theory-sweep

A Claude Code and Codex skill that asks many LLM "discipline experts" to read one design document and propose theories from their field that the design could borrow.

One investigator per discipline (philosophy, accounting, control engineering, music, … 52 fields by default) reads your design document and writes a short, fixed-format report of theory-integration candidates with sources. One model merges the reports by design group and by cross-cutting problem. Scripts and a judge model check format, citations and whether each candidate really connects to a sentence in the design.

Every result carries the same status line: an AI survey and comparison proposal, not adopted, effectiveness not proven. The skill finds candidates. It does not prove that any of them works.

What is in the repo

theory-sweep/            the skill folder (install this)
  SKILL.md               the procedure, steps 0–6
  references/            prompt templates, report format, field list, adapters, model notes, evidence
    strings/en.json      fixed report strings, English
    strings/ko.json      fixed report strings, Korean
  scripts/               render prompts, score format, check citations, count, compare models
README.md
LICENSE

Install

git clone https://github.com/BUGGIEEEEE/theory-sweep theory-sweep-skill
# Claude Code
ln -s "$(pwd)/theory-sweep-skill/theory-sweep" ~/.claude/skills/theory-sweep
# Codex
ln -s "$(pwd)/theory-sweep-skill/theory-sweep" ~/.codex/skills/theory-sweep

Requirements: Python 3 (standard library only) and network access to Crossref, plus OpenAlex if you set a key. Skill folder locations can differ between tool versions; check your tool's documentation if the skill does not appear.

Quick start

Ask your agent for a theory sweep, for example: "Run a theory sweep on docs/design.md." The skill then asks for these inputs in step 0.

Input	Required	Example
Target documents and line ranges	yes