world-model-skills
CLAUDE.md4 starsChanged 5 months agoMITView on GitHub
World Model Skills is a reusable skill package for Claude Code agents doing world model research. It covers five method families — Dreamer, JEPA, MuZero, Diffusion World Models, Transformer World Models — and provides a structured skill definition (world-model-research/SKILL.md), metadata indexes, validation/build scripts, and an evaluation benchmark suite. Current version: v1.0.0-alpha — 185 papers, 40-prompt eval at 100% across all 6 metrics. All commands run from repo root. Only Python stdlib is required (3.11+). Output: corpus/code/extracted/{code_id}/ with STRUCTURE.txt tree overview.
- Installs packages
In this repositoryOne other file it gives its agents.
- world-model-researchSkill
The file
We split this CLAUDE.md into the rules it gives an agent
2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 13
All commands run from repo root.
- Distinctive: line 177
When answering, always situate work within its family:
≈ “Common” is an estimate: at least 1 in 5 of the 27,641 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
CLAUDE.md
This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
Project overview
World Model Skills is a reusable skill package for Claude Code agents doing world model research. It covers five method families — Dreamer, JEPA, MuZero, Diffusion World Models, Transformer World Models — and provides a structured skill definition (world-model-research/SKILL.md), metadata indexes, validation/build scripts, and an evaluation benchmark suite.
Current version: v1.0.0-alpha — 185 papers, 40-prompt eval at 100% across all 6 metrics.
Common commands
All commands run from repo root. Only Python stdlib is required (3.11+).
Metadata build and validation
# Build paper catalog from source index
python scripts/build_paper_catalog.py --input metadata/papers_index.csv --output metadata/paper_catalog.csv
# Validate evidence index
python scripts/validate_evidence_index.py --input metadata/evidence_index.csv --check-paths --json eval/results/evidence_index_validation.json
# Check metadata completeness for release gating
python scripts/check_metadata_completeness.py --input metadata/paper_catalog.csv --json eval/results/metadata_completeness_report.json# Extract code archives for agent-readable source access
python scripts/extract_code.py --all
python scripts/extract_code.py --code ijepa_mainOutput: corpus/code/extracted/{code_id}/ with STRUCTURE.txt tree overview.
Storage architecture
corpus/
├── papers/
│ ├── pdf/{family}/ # Original PDFs by method family
│ └── md/ # Converted markdown (flat, paper_id.md + paper_id_brief.md)
├── code/
│ ├── archives/{family}/ # Original zip archives by method family
│ └── extracted/{code_id}/ # Extracted source code + STRUCTURE.txt
└── _duplicates/ # Dedup stagingAll corpus/ content is gitignored. Paths in metadata/*.csv reference these locations.
PDF → Markdown conversion
# Convert all PDFs to structured markdown (requires docling)
python scripts/convert_pdfs.py --all --update-evidence
# Convert a specific paper
python scripts/convert_pdfs.py --paper dreamer_v3 --update-evidence
# Dry run to see what would be converted
python scripts/convert_pdfs.py --all --dry-runOutput: corpus/papers/md/{paper_id}.md (full) + {paper_id}_brief.md (abstract+conclusion).
Corpus ingestion pipeline
# Search arXiv and auto-download papers
python scripts/ingest_arxiv.py --query "world model reinforcement learning" --max 30
python scripts/ingest_arxiv.py --ids 2301.04104,1803.10122 --tier core
python scripts/ingest_arxiv.py --query "dreamer world model" --max 50 --resume
# Manually register non-arXiv papers
python scripts/ingest_manual.py --title "World Models" --year 2018 --authors "Ha, Schmidhuber" --tier core --file paper.pdf --method-family transformer
# Link papers to GitHub repositories
python scripts/link_github.py --all-missing --min-stars 10
python scripts/link_github.py --paper dreamer_v3These scripts require pip install -r requirements.txt (only requests needed).
Citation candidate generation
python world-model-research/scripts/make_citation_candidates.py --metadata metadata/paper_catalog.csv --query "dreamer,atari,planning" --topk 20 --output eval/results/candidates.csvEval workflow
# Extract one mode's prompts
python eval/filter_eval_by_mode.py --input eval/test_prompts_v1.jsonl --mode research_qa --output eval/results/qa_prompts.jsonl
# Summarize judged results
python eval/summarize_eval.py --input eval/results/<scored>.jsonl --output eval/results/<summary>.md
# Check evidence/mapping coverage
python eval/check_evidence_coverage.py --responses eval/results/<responses>.md --output eval/results/<coverage>.jsonArchitecture
Three-layer design (same as JEPA-skills)
- Skill layer (
world-model-research/) — 4 modes, evidence-binding contract, method family taxonomy - Metadata layer (
metadata/) — papers_index, paper_catalog, evidence_index, code_index, benchmark_index (new) - Eval layer (
eval/) — prompts, rubric (6 dimensions incl. benchmark_accuracy), results
What's different from JEPA-skills
| Feature | JEPA-skills | world-model-skills |
| Method families | 1 (JEPA) | 5 (Dreamer, JEPA, MuZero, Diffusion, Transformer) |
| Eval dimensions | 5 | 6 (+ benchmark_accurate) |
| Metadata fields | modality, task, evidence_level | + method_family, benchmark_family |
| Benchmark index | N/A | metadata/benchmark_index.csv |
| Min evidence (QA) | 2 paper | 3 paper |
| Min evidence (repro) | 1 paper + 1 code | 1 paper + 1 code + 1 benchmark |
| MVP paper count | 148 | 185 |
Eval benchmark reference (v1.0.0-alpha)
40 prompts (10 per mode), all metrics at 100%:
| Metric | Score | N |
| Factual correctness | 100.0% | 40 |
| Citation hit | 100.0% | 40 |
| Executability | 100.0% | 10 |
| Debug hit | 100.0% | 10 |
| Benchmark accuracy | 100.0% | 5 |
| Hallucination rate | 0.0% | 40 |
Results in eval/results/v1_0_0_baseline_scored.jsonl and v1_0_0_baseline_summary.md.
Current state
Papers
| Family | Count | Code repos |
| Dreamer | 55 | 12 |
| JEPA | 76 | 4 |
| MuZero | 20 | 1 |
| Diffusion | 16 | 3 |
| Transformer | 18 | 5 |
| Total | 185 | 25 |
Data assets
- 185 papers indexed in
papers_index.csvandpaper_catalog.csv - 0 unclassified — all 185 papers have method_family, modality, task, evidence_level
- 25 confirmed code repositories (13.5% coverage)
- 200+ papers converted to markdown via docling (full + brief versions)
- 8 benchmarks in
benchmark_index.csv - 17 evidence rows in
evidence_index.csv - 6 local code archives in
code_index.csv
Key papers covered
- Dreamer lineage: PlaNet, DreamerV1/V2/V3, DayDreamer, TD-MPC2, Dreamer-CDP
- MuZero lineage: MuZero, EfficientZero, Sampled MuZero, VQ Planning
- JEPA lineage: I-JEPA, V-JEPA-2, MC-JEPA, Var-JEPA, LeJEPA, ThinkJEPA
- Diffusion: DIAMOND (NeurIPS 2024), AVID, Nano World Models
- Transformer: IRIS (ICLR 2023), TECO, Decision Transformer
- Benchmarks: Atari 100k (SimPLe), DMControl, Crafter, Meta-World
- All canonical papers: 28/33 (85%)
- Action/Claim list with IDs
- Evidence table with source type and pointer
- Mapping line:
C# -> E#orA# -> E# - Confidence:
high|medium|low
Minimum evidence: research_qa ≥ 3 papers; reproduction_plan ≥ 1 paper + 1 code + 1 benchmark; debug_triage ≥ 1 code; writing_citation every claim → at least one paper.
Method family taxonomy
When answering, always situate work within its family:
| Family | Core | Key signal |
| Dreamer | RSSM + actor-critic in latent space | "dreamer", "rssm", "imagination" |
| JEPA | Predictive latent, no decoder | "jepa", "joint embedding predictive" |
| MuZero | MCTS + learned dynamics | "muzero", "mcts", "tree search" |
| Diffusion WM | Diffusion-based dynamics prediction | "diffusion" + "world model" |
| Transformer WM | Autoregressive transformer over latents | "transformer" + "world model" + "tokenizer" |
Relationship to JEPA-skills
Independent sibling repo. JEPA-skills covers the JEPA family in detail; world-model-skills references JEPA work as one of five families without duplicating metadata.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

