Add PrathamLearnsToCode/paper2code

agents/PrathamLearnsToCode__paper2code/README.md

@@ -0,0 +1,78 @@
+# paper2code
+
+An AI agent skill that converts any arxiv paper into a minimal, citation-anchored Python implementation — faithful to what the paper actually says, honest about everything it doesn't.
+
+## What It Does
+
+Give it an arxiv URL or paper ID and it will:
+
+1. **Fetch and parse** the PDF, treating appendices and figure captions as first-class sources
+2. **Identify the core contribution** — the one thing the paper claims to introduce
+3. **Audit every implementation detail** — classifying each as `SPECIFIED`, `PARTIALLY_SPECIFIED`, or `UNSPECIFIED` before writing a single line of code
+4. **Generate citation-anchored code** where every non-trivial decision links back to a specific paper section and equation (`# §3.2, Eq. 4`)
+5. **Produce a walkthrough notebook** with CPU-runnable sanity checks that quote the paper and show the corresponding code
+
+## Try It
+
+```bash
+# Install the skill
+npx skills add PrathamLearnsToCode/paper2code/skills/paper2code
+
+# Run in your agent
+/paper2code https://arxiv.org/abs/1706.03762
+
+# With options
+/paper2code 2006.11239 --framework jax --mode full
+```
+
+## What Makes It Different
+
+Most code generation tools fill ambiguous gaps silently and confidently. paper2code refuses to. If the paper doesn't specify a hyperparameter, the generated code marks it:
+
+```python
+# [UNSPECIFIED] Paper does not state epsilon for LayerNorm — using 1e-6 (common default)
+# Alternatives: 1e-5 (PyTorch default), 1e-8 (some implementations)
+self.norm = nn.LayerNorm(d_model, eps=1e-6)
+```
+
+Every specified detail is traced back to the paper:
+
+```python
+# §3.2, Eq. 2 — attention_weights = softmax(QK^T / sqrt(d_k))
+attn_out = self.attention(self.norm1(x))
+```
+
+## Output Structure
+
+```
+attention_is_all_you_need/
+├── README.md                    # Paper summary and quick-start
+├── REPRODUCTION_NOTES.md        # Ambiguity audit — every unspecified choice documented
+├── requirements.txt             # Pinned dependencies
+├── src/
+│   ├── model.py                 # Architecture with section citations
+│   ├── loss.py                  # Loss functions with equation references
+│   ├── data.py                  # Dataset skeleton with preprocessing TODOs
+│   ├── train.py                 # Training loop (full/educational mode)
+│   └── evaluate.py              # Metric computation
+├── configs/
+│   └── base.yaml                # All hyperparams — cited or flagged [UNSPECIFIED]
+└── notebooks/
+    └── walkthrough.ipynb        # Runnable on CPU, quotes paper passages
+```
+
+## Modes
+
+| Mode | Description |
+|------|-------------|
+| `minimal` (default) | Core contribution only |
+| `full` | Core + training loop + data pipeline + evaluation |
+| `educational` | Adds extra inline comments, theory sections in notebook |
+
+## Worked Examples
+
+The repo ships with fully worked examples for:
+- **Attention Is All You Need** (`1706.03762`) — the original Transformer
+- **DDPM** (`2006.11239`) — denoising diffusion probabilistic models
+
+Each includes the complete generated output plus an honest `review.md` evaluating what was right and where gaps were flagged.

agents/PrathamLearnsToCode__paper2code/metadata.json

@@ -0,0 +1,14 @@
+{
+  "name": "paper2code",
+  "author": "PrathamLearnsToCode",
+  "description": "Converts any arxiv paper into a citation-anchored Python implementation — audits ambiguity, never silently fills gaps, cites every code decision back to the paper.",
+  "repository": "https://github.com/PrathamLearnsToCode/paper2code",
+  "version": "1.0.0",
+  "category": "research",
+  "tags": ["arxiv", "machine-learning", "paper-implementation", "citation-anchored", "research", "pytorch"],
+  "license": "MIT",
+  "model": "claude-sonnet-4-6",
+  "adapters": ["claude-code", "system-prompt"],
+  "icon": false,
+  "banner": false
+}