Skip to content

CLAUDE.md

Latest commit

 

History

History
58 lines (39 loc) · 3.35 KB

CLAUDE.md

File metadata and controls

58 lines (39 loc) · 3.35 KB

Skillception

This repo is an experiment harness, not a library or application. It tests how many layers of meta-recursive skill creation Claude can maintain.

Tone

This is a humor project. The README is written in a dry, self-aware comedic voice — treat the whole thing as a fun experiment that happens to be rigorous, not a serious benchmark that happens to be funny. Match that tone in any writing (docs, commit messages, comments). Keep it light.

What this repo does

The harness runs a recursive loop where each step calls claude -p twice: once for an executor (generates a SKILL.md) and once for a judge (blindly evaluates what meta-level that SKILL.md targets). A run proceeds in rounds of ascending meta-level then cascading back down, stopping on the first mismatch between expected and detected level. See .context/nomenclature.md for the canonical definitions of levels, rounds, and naming conventions.

Website

Live at https://skillception.study. Source in website/ — React/TypeScript/Vite/Tailwind. Displays experiment results dynamically from runs/ data.

Website commands (from repo root)

npm run website:dev          # Vite dev server
npm run website:build        # TypeScript check + production build
npm run website:typecheck    # TypeScript only
npm run website:lint         # ESLint
npm run website:test         # Vitest
npm run website:export-data  # Regenerate website/public/results.json from runs/

Data flow

  1. scripts/export_results.py reads runs/*/result.json and writes website/public/results.json
  2. Run npm run website:export-data after new experiment runs to update the website data
  3. The React app fetches results.json at runtime and computes statistics client-side

Key files

  • agents/executor.md — prompt template interpolated into executor's claude -p call. Not a registered Claude Code agent — plain Markdown read by the Python harness.
  • agents/judge.md — prompt template for the judge. Same pattern.
  • scripts/run_experiment.py — main harness. Orchestrates claude -p subprocess calls, manages the round-based loop, logs results as JSON.
  • scripts/analyze_results.py — reads result JSONs from runs/, prints aggregate statistics.
  • scripts/export_results.py — exports cleaned result JSONs for the website.
  • website/ — React/TS/Vite/Tailwind results website.

Architecture decisions

  • Each claude -p call strips the CLAUDECODE env var to avoid the recursive-invocation guard.
  • The executor gets --allowedTools Write only. The harness pre-creates output directories.
  • The judge gets no tools — just text analysis returning JSON.
  • --max-turns 10 caps each call to prevent runaway loops.
  • --output-format json returns a {"result": "...", "usage": {...}} envelope.
  • Windows paths are converted to forward slashes before embedding in prompts.

Style

This repo has a visual identity defined in .claude/skills/skillception-style/SKILL.md. Consult it before creating any visual or written assets. The short version: everything is styled as a formal academic paper from a fictional department. The humor is in the contrast between the serious presentation and the absurd content. Never explain the joke.

Runtime directory (gitignored)

  • runs/ — one subdirectory per run, each containing result.json and skills/ with generated SKILL.md files organized by step