Paper2Code: AI Agent for Converting Research Papers into Code

Overview

This Skill executes a 4+2 stage pipeline effectively systematically analyzing research papers and converting them into executable code.

Core Principle: Do not simply read the paper and generate code; generate a structured intermediate representation (YAML) first, then write the code.

---

⚠️ Critical Behavioral Control Rules (CRITICAL)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚠️ MANDATORY BEHAVIORAL RULES
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

1. Implement one file at a time
2. Proceed to the next file only after completing the current file, without asking for confirmation
3. Original paper specifications always take precedence over reference code
4. Perform a Self-Check for each Phase before completion
5. Save all intermediate results as YAML files

DO:
✓ Implementing exactly what is stated in the paper
✓ Write simple and direct code
✓ Working code first, elegant code later
✓ Test each component immediately
✓ Move to the next file immediately after implementation is complete

DON'T:
✗ Do not ask "Shall I implement the next file?" between files
✗ Extensive documentation not required for core functionality
✗ Optimization not needed for reproducibility
✗ Excessive abstraction or design patterns
✗ Providing instructions without writing actual code
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

---

Input Processing

Supported Formats

1. arXiv URL:

   https://arxiv.org/abs/xxxx.xxxxx

   or

   https://arxiv.org/pdf/xxxx.xxxxx.pdf

2. PDF File Path:

   /path/to/paper.pdf

3. Converted Text/Markdown: When paper content is provided as text

Input Processing Method

For arXiv URL:

# Convert to PDF URL and download
curl -L "https://arxiv.org/pdf/xxxx.xxxxx.pdf" -o paper.pdf

# Convert PDF to text (using pdftotext)
pdftotext -layout paper.pdf paper.txt

For PDF File:

pdftotext -layout "/path/to/paper.pdf" paper.txt

---

Pipeline Overview

[User Input: Paper URL/File]
        │
        ▼
┌─────────────────────────────────────────────┐
│ Step 0: Acquire Paper Text                  │
│ - arXiv URL → Download PDF                  │
│ - PDF → Convert to Text                     │
└─────────────────────────────────────────────┘
        │
        ▼
┌─────────────────────────────────────────────┐
│ Phase 0: Search Reference Code (Optional)   │
│ @[05_reference_search.md]                   │
│ Output: reference_search.yaml               │
└─────────────────────────────────────────────┘
        │
        ▼
┌─────────────────────────────────────────────┐
│ Phase 1: Algorithm Extraction               │
│ @[01_algorithm_extraction.md]               │
│ Output: 01_algorithm_extraction.yaml        │
└─────────────────────────────────────────────┘
        │
        ▼
┌─────────────────────────────────────────────┐
│ Phase 2: Concept Analysis                   │
│ @[02_concept_analysis.md]                   │
│ Output: 02_concept_analysis.yaml            │
└─────────────────────────────────────────────┘
        │
        ▼
┌─────────────────────────────────────────────┐
│ Phase 3: Implementation Plan                │
│ @[03_code_planning.md]                      │
│ Output: 03_implementation_plan.yaml         │
└─────────────────────────────────────────────┘
        │
        ▼
┌─────────────────────────────────────────────┐
│ Phase 4: Code Implementation                │
│ @[04_implementation_guide.md]               │
│ Output: Complete Project Directory          │
└─────────────────────────────────────────────┘

---

Data Transfer Format Between Stages

Phase 1 → Phase 2 Transfer

phase1_to_phase2:
  algorithms_found: "[Number of found algorithms]"
  key_algorithms:
    - name: "[Algorithm Name]"
      section: "[Paper Section]"
      complexity: "[Simple/Medium/Complex]"
  hyperparameters_count: "[Number of collected hyperparameters]"
  critical_equations: "[List of critical equation numbers]"
  missing_info: "[List of missing information]"

Phase 2 → Phase 3 Transfer

phase2_to_phase3:
  components_count: "[Number of identified components]"
  implementation_complexity: "[Low/Medium/High]"
  key_dependencies:
    - "[Component A] → [Component B]"
  experiments_to_reproduce:
    - "[Experiment Name]: [Expected Result]"
  success_criteria:
    - "[Specific Success Criteria]"

Phase 3 → Phase 4 Transfer

phase3_to_phase4:
  file_order: "[List of files in implementation order]"
  current_file: "[Currently implementing file]"
  completed_files: "[List of completed files]"
  blocking_dependencies: "[Dependencies to resolve]"

---

Detail of Each Phase

Phase 0: Reference Code Search (Optional)

Using the @05_reference_search.md prompt:

• Search for and evaluate 5 similar implementations
• Secure references to improve implementation quality
• Output: Reference list in YAML format

Phase 1: Algorithm Extraction

Using the @01_algorithm_extraction.md prompt:

• Extract all algorithms, equations, and pseudocode
• Collect hyperparameters and configuration values
• Organize training procedures and optimization methods
• Output: Complete algorithm specification in YAML format

Phase 2: Concept Analysis

Using the @02_concept_analysis.md prompt:

• Map paper structure and sections
• Analyze system architecture
• Identify component relationships and data flow
• Organize experiment and validation requirements
• Output: Implementation requirements specification in YAML format

Phase 3: Establish Implementation Plan

Using the @03_code_planning.md prompt:

• Integrate results from Phase 1 and 2
• Generate detailed implementation plans for 5 essential sections:

  1. file_structure

     : Project file structure

  2. implementation_components

     : Implementation component details

  3. validation_approach

     : Validation and testing methods

  4. environment_setup

     : Environment and dependencies

  5. implementation_strategy

     : Step-by-step implementation strategy
• Output: Complete YAML implementation plan (8000-10000 characters)

Phase 4: Code Implementation

Following the guide @04_implementation_guide.md:

• Generate code file by file according to the plan
• Implement in dependency order
• Each file must be complete and executable
• Output: Executable codebase

---

Memory Management

Refer to the guide @06_memory_management.md:

• Context management when processing long papers
• Saving step-by-step outputs
• Recovery protocol in case of interruption

---

Quality Standards

Principles that Must Be Followed

• Completeness: Complete implementation without placeholders or TODOs
• Accuracy: Accurately reflect equations and parameters specified in the paper
• Executability: Code that can be executed immediately
• Reproducibility: Must be able to reproduce the results of the paper

File Implementation Order

1. Configuration and environment files (config, requirements.txt initialization)
2. Core utilities and base classes
3. Main algorithm/model implementation
4. Training and evaluation scripts
5. Documentation (README.md, requirements.txt finalization)

---

✅ Final Completion Checklist (MANDATORY)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚠️ BEFORE DECLARING COMPLETE - ALL MUST BE YES
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

□ All algorithms in the paper implemented?       → YES / NO
□ Correct versions of environment/datasets set?  → YES / NO
□ All comparison methods referenced implemented? → YES / NO
□ Working integration to run paper experiments?  → YES / NO
□ All metrics, figures, tables reproducible?     → YES / NO
□ Basic docs explaining how to reproduce?        → YES / NO
□ Code runs without errors?                      → YES / NO

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚠️ If even one is NO, it is NOT complete!
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

---

Usage Examples

Example 1: arXiv Paper

User: Implement this paper https://arxiv.org/abs/2301.12345

Claude: I will analyze the paper and convert it to code.

[Phase 0: Reference Code Search (Optional)...]
[Phase 1: Algorithm Extraction...]
[Phase 2: Concept Analysis...]
[Phase 3: Establish Implementation Plan...]
[Phase 4: Code Generation...]

Example 2: PDF File

User: Implement the algorithms from this paper /home/user/papers/attention.pdf

Example 3: Specific Request

User: Implement only the algorithm in Section 3 of this paper

---

Related Files

• 01_algorithm_extraction.md - Phase 1: Algorithm Extraction
• 02_concept_analysis.md - Phase 2: Concept Analysis
• 03_code_planning.md - Phase 3: Implementation Plan
• 04_implementation_guide.md - Phase 4: Implementation Guide
• 05_reference_search.md - Phase 0: Reference Search (Optional)
• 06_memory_management.md - Memory Management Guide

---

Precautions

⚠️ REMEMBER:

1. Read the paper thoroughly: Start implementation after understanding the entire content
2. Save detailed results: Save YAML output of each Phase as a file
3. Incremental implementation: Do not generate all code at once, proceed file by file
4. Include verification: Include simple test code if possible
5. Reference is inspiration: Reference code is for understanding and application, not copying