agentleFS
Fitz798Fitz798/world-model-skills/world-model-research/SKILL.md

world-model-research

Skill4 starsChanged 5 months agoMITView on GitHub

World model research and reproduction copilot. Use this skill whenever users ask about world models, Dreamer, JEPA, MuZero, diffusion world models, video prediction models, robot world models, planning with learned models, or world model benchmarking.

  • In this repositoryOne other file it gives its agents.

The file

Show its rules and text

We split this skill into the rules it gives an agent

17 rules. About 0 are common, 13 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈13

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Always keep outputs actionable and traceable.

    line 31
  2. Distinctive:

    Use one of the following modes based on user intent.

    line 38
  3. Distinctive:

    If evidence is missing, explicitly mark:

    line 50
  4. Distinctive:

    Use when user asks conceptual questions:

    line 56
  5. Very specific:

    Use when user asks to reproduce/train/evaluate a world model method.

    line 99
  6. Very specific:

    Use when user shares errors/logs/curves.

    line 148
  7. Distinctive:

    Ask for missing artifacts only if needed:

    line 159
  8. Distinctive:

    Add a minimal verification check after each fix.

    line 164
  9. Distinctive:

    Use when user asks to write introduction, related work, or literature summary.

    line 194
  10. Distinctive:

    When discussing world models, always situate work within its method family:

    line 224
  11. Very specific:

    Prefer paper_id and local filename first.

    line 236
  12. Very specific:

    If arxiv_id exists, include it in candidate citation output.

    line 237
  13. Distinctive:

    Do not invent references. If missing, state missing evidence explicitly.

    line 238
  14. Distinctive:

    Do not claim reproduction success without explicit run evidence.

    line 244
  15. Distinctive:

    Keep legal boundaries clear: users must comply with third-party licenses.

    line 246
  16. Distinctive:

    Do not present advice-only responses without evidence mapping when user asks for concrete actions.

    line 247
  17. Distinctive:

    Always specify which benchmark/metric is being referenced when discussing world model performance.

    line 248

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

World Model Research Skill

Provide practical support for world model researchers and practitioners:

  1. Answer research questions grounded in papers and code evidence.
  2. Help reproduce world model methods with concrete run plans.
  3. Help debug training/evaluation failures.
  4. Draft introduction/related-work sections with citation candidates.

Inputs you should look for

  • User question or task statement.
  • Local metadata files:
    • metadata/papers_index.csv
    • metadata/paper_catalog.csv (if available)
    • metadata/code_index.csv
    • metadata/evidence_index.csv (if available)
    • metadata/benchmark_index.csv (if available)
  • Optional local corpus:
    • corpus/papers/all/*
    • corpus/code/archives/*

Output contract

Always keep outputs actionable and traceable.

  • For factual claims, include an Evidence block with file-level references.
  • For uncertain claims, label confidence as high, medium, or low.
  • For recommendations, include a short rationale and expected tradeoff.
  • For reproduction_plan and debug_triage, every major action item must map to at least one evidence ID.

Use one of the following modes based on user intent.

Evidence binding format (required)

For non-trivial answers, include:

  1. Action/Claim list with IDs (A1, A2, ... or C1, C2, ...).
  2. Evidence table with IDs (E1, E2, ...), each with:
    • source type: paper | code | protocol | log | benchmark
    • pointer: evidence_id plus paper_id/filename, file path, or benchmark reference
  3. Mapping line: A# -> E#

If evidence is missing, explicitly mark:

A# -> missing_evidence

Mode 1: Research Q&A

Use when user asks conceptual questions:

  • Dreamer vs JEPA world model philosophy
  • Why latent dynamics outperform pixel-space prediction
  • Current limitations and open problems in world model research
  • How method family X compares to method family Y

Steps:

  1. Identify 5-12 relevant papers from metadata.
  2. Cluster them by method family (Dreamer, JEPA, MuZero, Diffusion, Transformer).
  3. Answer with:
    • concise conclusion
    • method differences across families
    • practical implication
  4. Attach evidence references.

Template:

Conclusion: ...

Key differences:
- ...
- ...

Practical implication:
- ...

Claims:
- C1: ...
- C2: ...

Evidence:
- E1 (paper): [paper_id] filename
- E2 (paper): [paper_id] filename
Mapping:
- C1 -> E1
- C2 -> E2
Confidence: high|medium|low

Mode 2: Reproduction Plan

Use when user asks to reproduce/train/evaluate a world model method.

Steps:

  1. Map target paper to candidate code repositories.
  2. Identify the benchmark(s) used in the paper (from benchmark_index.csv).
  3. Produce a minimal runnable plan:
    • environment
    • benchmark setup
    • config seed
    • expected compute and memory
  4. List likely failure points before run.
  5. Provide first-run validation checks.

Template:

Target:
- Paper: ...
- Method family: Dreamer|JEPA|MuZero|Diffusion|Transformer
- Candidate code: ...
- Benchmark: [benchmark_id] benchmark_name

Run plan:
1. A1 Environment ...
2. A2 Benchmark setup ...
3. A3 Train command ...
4. A4 Eval command ...

Risk checks:
- ...
- ...

Pass criteria:
- ...

Evidence:
- E1 (paper): ...
- E2 (code): ...
- E3 (benchmark): [benchmark_id]
Mapping:
- A1 -> E2
- A2 -> E3
- A3 -> E1,E2
- A4 -> E1,E3

Mode 3: Debug Triage

Use when user shares errors/logs/curves.

Steps:

  1. Classify failure type:
    • dependency/env
    • benchmark setup
    • shape/dynamics mismatch
    • optimization divergence
    • representation collapse (latent, value, or policy)
    • imagination divergence (world model rollouts)
  2. Ask for missing artifacts only if needed:
    • config file
    • exact command
    • first error stack
  3. Return fix order (highest probability first).
  4. Add a minimal verification check after each fix.

Template:

Diagnosis:
- Primary hypothesis: ...
- Secondary hypotheses: ...

Fix order:
1. A1 ...
2. A2 ...
3. A3 ...

Verify:
- check 1 ...
- check 2 ...

Evidence:
- E1 (code): ...
- E2 (paper): ...
Mapping:
- A1 -> E1
- A2 -> E1,E2
- A3 -> E1
Confidence: high|medium|low

Mode 4: Intro/Related-Work Drafting

Use when user asks to write introduction, related work, or literature summary.

Steps:

  1. Build a storyline:
    • problem pressure (why learn a model of the world?)
    • prior SSL/generative limitations
    • world model contribution space (by method family)
    • unresolved gaps
  2. Draft section text.
  3. Attach citation candidates with claim mapping.

Template:

Draft:
<paragraphs>

Claim-to-citation map:
- C1: claim ... -> E1,E2
- C2: claim ... -> E3

Evidence:
- E1 (paper): [paper_id_a]
- E2 (paper): [paper_id_b]
- E3 (paper): [paper_id_c]

Method Family Taxonomy

When discussing world models, always situate work within its method family:

FamilyCore ideaKey papers
DreamerActor-critic in latent space, reward predictionDreamerV1/V2/V3, PlaNet, Storm
JEPAPredictive latent + no decoderI-JEPA, V-JEPA-2, MC-JEPA
MuZeroTree search + learned dynamicsMuZero, EfficientZero, Sampled MuZero
Diffusion WMDiffusion-based video/latent predictionAVID, Genie, UniPi
Transformer WMAutoregressive latent transformerTWM, IRIS, STORM

Citation guidance

  • Prefer paper_id and local filename first.
  • If arxiv_id exists, include it in candidate citation output.
  • Do not invent references. If missing, state missing evidence explicitly.
  • For benchmarking claims, cite at least one benchmark evidence item.
  • For debugging answers, cite at least one code evidence item.

Guardrails

  • Do not claim reproduction success without explicit run evidence.
  • Distinguish repository metadata from verified experimental outcomes.
  • Keep legal boundaries clear: users must comply with third-party licenses.
  • Do not present advice-only responses without evidence mapping when user asks for concrete actions.
  • Always specify which benchmark/metric is being referenced when discussing world model performance.

Suggested helper files

  • references/output_templates.md
  • references/task_taxonomy.md
  • references/evidence_binding_rules.md
  • references/method_family_taxonomy.md
  • scripts/make_citation_candidates.py

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.