SKILL.md
Paper Implementation
Input
Require arxiv link (not PDF). Fetch LaTeX: https://arxiv.org/e-print/XXXX.XXXXX (.tar.gz with .tex files)
Phases
1. Discovery
Parse LaTeX for: architecture, algorithms, hyperparameters, loss functions, datasets. Search existing GitHub implementations. Note ambiguities.
Gather: Scope (train/inference/finetune)? Scale (model size, compute)? Baseline codebase? Priority (accuracy/speed/memory)? Validation method?
2. Verification
If repo exists: audit against LaTeX source. Check architecture, hyperparameters, training procedure. Identify discrepancies.
3. Refinement
Present findings, ask questions, iterate until execution steps are perfectly clear.
4. Implementation
Build/modify code. Write correctness tests. Profile performance.
5. Optimization (optional)
Profile with nsight-systems/torch.profiler. Write custom CUDA/Triton kernels. Benchmark with measurements.
Compute Strategy
- Local 3090: quick tests, debugging, small validation
- VP H100s: training runs, large experiments (ask before provisioning)
On arxiv link
- Parse LaTeX source
- Search existing implementations
- Present structured questions
- Wait for answers, refine, then proceed autonomously