Paper To Skill

Paper-to-Skill Pipeline

Transform research papers into production-grade skill packages. The pipeline extracts the actionable methodology from a paper, structures it as a skill specification, and feeds it through co-evolutionary refinement to produce a validated package.

This closes the loop between research and practice: a paper published today can become an executable skill tomorrow, without manual authoring.

Reference Files

| File | Contents | Load When | | ------------------------------------- | ------------------------------------------------- | --------- | | references/extraction-patterns.md | Patterns for extracting methodology from papers | Always |

Prerequisites

• The to-markdown skill (for PDF/document conversion)
• The research-critique skill (for paper analysis)
• The test-engineer agent (for co-evolutionary skill generation)

Workflow

Phase 1: Paper Intake

Accept the paper in any supported format:

| Input Format | Action | | --------------------- | ---------------------------------------------------------- | | arXiv ID (e.g., 2604.01687) | Fetch via https://arxiv.org/abs/, convert PDF | | arXiv URL | Extract ID, fetch and convert | | PDF file path | Convert using to-markdown skill | | URL to paper | Fetch via WebFetch, convert if PDF | | Pasted text | Use directly |

For PDF conversion, invoke the to-markdown skill: > Convert this PDF to clean markdown, preserving section structure, tables, equations, > and algorithm pseudocode. Drop references section but keep inline citations.

Phase 2: Critical Analysis

Invoke the research-critique skill on the converted paper:

> Analyze this paper focusing on: > 1. Core contribution: what is the novel methodology? > 2. Algorithm description: extract the step-by-step procedure > 3. Input/output specification: what goes in, what comes out? > 4. Key parameters and their valid ranges > 5. Claimed results and the evidence supporting them > 6. Failure modes and limitations acknowledged by the authors > 7. Prerequisites and dependencies (tools, data, compute)

The critique output becomes the foundation for the skill specification.

Phase 3: Skill Specification Extraction

From the critique output, build a structured skill specification:

specification:
  name: 
  domain: 
  source_paper:
    title: 
    arxiv_id: 
    url: 
    authors: 
    date: 
  
  capabilities:
    - 
    - 
    - 
  
  input_format: 
  output_format: 
  
  algorithm_steps:
    - step: 1
      description: 
      parameters: []
    - step: 2
      description: 
  
  failure_modes:
    - 
  
  example_tasks:
    - 
    - 
    - 

Extraction rules:

• Prefer the paper's own algorithm pseudocode over prose descriptions
• Include parameter ranges from the paper's experiments (e.g., "learning rate: 0.001-0.01")
• Map the paper's terminology to armory conventions (e.g., "module" → "skill", "pipeline" → "workflow")
• If the paper describes multiple variants, extract the best-performing one

See references/extraction-patterns.md for patterns specific to common paper types.

Phase 4: Skill Generation

Hand off the specification to the test-engineer agent for co-evolutionary generation:

> Evolve a skill for: [specification.domain] > > Capabilities: [specification.capabilities] > Algorithm: [specification.algorithmsteps] > Input: [specification.inputformat] > Output: [specification.outputformat] > Failure modes: [specification.failuremodes] > Example tasks: [specification.exampletasks] > > Source: [specification.sourcepaper.title] ([specification.source_paper.url])

The test-engineer runs its full co-evolutionary loop (generate → verify → oracle → refine) using the specification as the task description.

Phase 5: Attribution and Finalization

Ensure the generated skill properly attributes the source paper:

1. Frontmatter: Add source: to the metadata
2. Body: Include an attribution section at the end of SKILL.md:

```markdown ## Attribution

This skill implements the methodology from: > > > > ```

1. References: If the paper has supplementary materials (code, datasets), create a

source materials reference file in the generated skill's references/ directory linking to them

1. Verify the skill name does not conflict with existing packages in manifest.yaml

Output

The complete skill package at skills//:

• SKILL.md with attribution and paper-derived workflow
• evals/cases.yaml with assertions generated by the co-evolutionary loop
• references/ with extraction patterns and source materials
• evals/evolution-log.yaml from the test-engineer's refinement process

Error Handling

| Error | Resolution | | ------------------------------------ | --------------------------------------------------------- | | Paper has no clear algorithm | Extract the methodology from the experiments section | | Paper is purely theoretical | Report: no actionable methodology; suggest literature-review instead | | PDF conversion fails | Try alternative: fetch HTML version or request user paste text | | Paper methodology requires data/compute | Note in skill's prerequisites; skill may be a workflow template only | | test-engineer budget exhausted | Return best-scoring iteration with manual review warning |

Limitations

• Cannot extract visual methodologies (circuit diagrams, neural architecture figures)

— works on textual algorithm descriptions only

• Papers with multiple interdependent contributions may produce overly complex skills

— consider splitting into multiple skills

• Non-English papers require translation before processing
• The generated skill's quality depends on the paper's clarity of methodology description

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Mathews-Tom
• Source: Mathews-Tom/armory
• License: MIT
• Homepage: https://mathews-tom.github.io/armory/

Install and usage instructions live in the source repository linked above.