Paper First Principles

Overview

Convert academic papers into progressive, engineer-friendly documentation using first principles thinking.

When to Use

Use when:

• You need to understand a research paper deeply from an engineering perspective
• Extracting design patterns, system architectures, or algorithms from academic literature
• Preparing technical knowledge sharing or documentation for engineering teams
• Mapping theoretical research to practical software engineering applications

Don't use for:

• Papers without clear technical contributions (survey papers, opinion pieces)
• One-off reading without need for structured documentation
• Standard practices already well-documented elsewhere

Quick Start

# Basic usage
kimi paper-first-principles https://arxiv.org/abs/xxxx.xxxxx

# Engineer perspective with domain focus
kimi paper-first-principles paper.pdf --audience engineer --domain distributed-systems

# Output to file
kimi paper-first-principles paper.pdf --output ./docs/analysis.md

Output Structure

Generated documents contain 8 standard sections:

Paper Types & Progressive Paths

Parameters

--audience

• engineer: Implementation details, design patterns, edge cases
• researcher: Technical depth, related work comparison, theory
• manager: Problem context, decision rationale, risk assessment

--domain (optional)

Engineering domain for contextual mapping:

• distributed-systems: Distributed systems, microservices
• storage: Storage systems, file systems
• database: Databases, data warehouses
• network: Networks, CDN, load balancing
• ml-system: ML systems, recommendation systems

Analysis Workflow

This skill processes papers in 6 stages using prompts in prompts/:

1. Core Extraction (extract_core.txt): Identify contributions and key decisions
2. First Principles (first_principles.txt): Trace problem essence and rationale
3. Progressive Layers (progressive_layers.txt): Organize by complexity
4. Engineering Map (engineering_map.txt): Map to software patterns (engineer audience only)

Output Templates

Templates in templates/ provide structure for each paper type:

• system.md: System papers (infrastructure, architecture)
• algorithm.md: Algorithm papers (models, methods)
• theory.md: Theory papers (analysis, proofs)

Examples

See examples/attention_residuals.md for a complete example converting the Attention Residuals paper into an engineer-friendly analysis with distributed systems mappings.

Resource Loading Guide

Always load in this order:

1. Parse paper → Extract core (use prompts/extract_core.txt)
2. First principles analysis (use prompts/first_principles.txt)
3. Progressive organization (use prompts/progressive_layers.txt)
4. Engineering mapping (if --audience engineer, use prompts/engineering_map.txt)
5. Generate output (use appropriate template from templates/)

Constraints & Notes

• Paper quality matters: clear abstract and introduction required
• Output depth auto-adjusts based on --audience
• Engineering mapping accuracy depends on --domain setting
• Complex proofs may need manual verification