---

name: papers-in-100-lines-of-code description: "Routes Papers-in-100-Lines-of-Code catalog lookup, compact ML paper implementation adaptation, dependency planning, backend safety, and troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: MIT

Papers-in-100-Lines-of-Code

Use this repo skill when the user asks about the Papers-in-100-Lines-of-Code repository, compact educational implementations of ML papers, or safe planning for adapting/running those implementations. The source is a catalog of 62 standalone paper mini-projects, not one installable Python package.

Start here

• Read repo provenance before checking whether this skill matches a current checkout or before refreshing it.
• Read implementation index, or query implementation-index.json with query_implementation_index.py, when the task names a paper, method, script, symbol, or family.
• Read dependency and backend guide before installing requirements or attempting a full run.
• Read repo-level troubleshooting for installability, dependency conflicts, data/weight downloads, CUDA, and output side effects.
• Run check_skill_assets.py when validating the generated skill tree itself.

Route by user intent

User task	Route
Find whether a paper is implemented, choose an entry, interpret requirements, or plan a safe first run	paper-catalog-and-execution
Adapt or explain GANs, VAEs, normalizing flows, diffusion/DDIM/PNDM/DPM-Solver, DreamBooth, Stable Diffusion, image translation, SNL, or AALR	generative-models
Work with NeRF, Fourier/SIREN/MFN implicit fields, PlenOctrees/Plenoxels/K-Planes, 3D Gaussian Splatting, Speedy-Splat, Spherical Voronoi, Splatter Image, cameras, rays, or rendering memory	neural-rendering-3d
Adapt optimizers, activations, layers, MAML/Reptile/hypergradients, Deep Image Prior, DQN/DDQN/PPO, or long-loop training examples	optimization-meta-rl

Operating posture

• Do not install every requirements file into one environment. Choose one paper or compatible family first, then use its per-entry pins.
• Do not treat full upstream scripts as quick smoke tests. Many train for thousands to millions of steps, write output images, hard-code CUDA, or need external datasets/weights.
• Do not report full paper reproduction unless the user supplied the data, weights, hardware/backend, output policy, and budget and the run actually passed.
• For algorithm understanding, prefer tiny offline adaptations with explicit device, data, and output arguments.
• For CUDA-heavy rendering/text-to-image/RL tasks, verify the framework/backend separately before starting a long run.

Minimal generated-skill checks

From this skill directory, safe checks are stdlib-only:

python scripts/query_implementation_index.py --query "stable diffusion"
python scripts/query_implementation_index.py --group neural-rendering-3d --limit 5
python scripts/check_skill_assets.py

These checks validate the generated catalog and routing assets. They do not execute or reproduce upstream paper implementations.

When not to use this skill

• Do not use it as a replacement for a specialized production library skill such as Diffusers, PyTorch, Gym, or a full NeRF framework when the user is not working with this paper-code catalog.
• Do not use it to claim benchmark parity with any paper. The repository favors compact educational clarity over exhaustive reproduction infrastructure.
• Do not use it for ordinary repository maintenance unless the change affects the paper catalog, per-paper requirements, or implementation scripts.

Scope limits

This skill distills source evidence into operating guidance and bundled helper scripts. It does not bundle every full upstream training/rendering program, does not fetch datasets or weights, and does not verify optional full CUDA paper runs. When a user needs a full reproduction, use the catalog and owning sub-skill to plan an isolated environment, data/weight acquisition, runtime budget, and backend validation first.