A reusable skill package for Claude Code, Codex, and similar coding/research agents focused on world model research workflows. Covers five method families: Dreamer, JEPA, MuZero, Diffusion World Models, and Transformer World Models.
Current version: v1.0.0-alpha — 187 papers, 20-prompt eval at 100% across all metrics.
| Metric | Score |
|---|---|
| Factual correctness | 100.0% |
| Citation hit | 100.0% |
| Executability | 100.0% |
| Debug hit | 100.0% |
| Benchmark accuracy | 100.0% |
| Hallucination rate | 0.0% |
world-model-research/SKILL.md: World model research skill definition (4 modes, method family taxonomy).metadata/: structured indexes for papers, code, evidence, and benchmarks.scripts/: metadata build, evidence validation, and completeness checking.eval/: multi-mode eval suite with scoring rubric.docs/: roadmap, evidence schema, paper-code linking policy, release workflow.CLAUDE.md: project rules for Claude Code agents working in this repo.
- Full paper PDFs and full third-party codebases should usually stay local.
- For open-source publishing, share metadata + scripts; fetch source artifacts on demand.
.
├── world-model-research/
│ ├── SKILL.md
│ ├── references/
│ └── scripts/
├── metadata/
├── scripts/
├── eval/
└── corpus/ # local cache; large files ignored by default
| Family | Core idea | Representative papers |
|---|---|---|
| Dreamer | RSSM + actor-critic in latent space | DreamerV1, DreamerV2, DreamerV3 |
| JEPA | Predictive latent, no decoder | I-JEPA, V-JEPA-2, Var-JEPA |
| MuZero | MCTS + learned dynamics | MuZero, EfficientZero |
| Diffusion WM | Diffusion-based dynamics prediction | AVID, UniPi |
| Transformer WM | Autoregressive transformer over latents | TWM, IRIS, STORM |
- Put your local paper/code assets under
corpus/(optional for public repo users). - Build the paper catalog:
python scripts/build_paper_catalog.py --input metadata/papers_index.csv --output metadata/paper_catalog.csv- Validate data integrity:
python scripts/validate_evidence_index.py --input metadata/evidence_index.csv
python scripts/check_metadata_completeness.py --input metadata/paper_catalog.csv- Use the skill — ask about world model methods, reproduction, debugging, or drafting.
- Python stdlib only. No pip dependencies.
- Python 3.11+. Tested with Python 3.13.5.
- Run commands from repo root.
This repo is an independent sibling of JEPA-skills. JEPA-skills covers the JEPA method family in detail; world-model-skills references JEPA as one of five families. The two repos share architecture but maintain separate metadata and eval suites.
MIT. Users are responsible for third-party paper/code licenses.