research-ingestion

Research Ingestion

Analyze research papers and extract actionable insights for OpenClaw agent workflows.

Trigger Phrases

• "analyze this paper", "read this research", "extract insights from"
• "turn this paper into a skill", "what can we learn from this paper"
• "research to skill", "/research"

Workflow

1. Obtain Paper Text

From URL (arXiv, PDF link, blog post):

# Use the summarize skill for URLs
# Or fetch + extract directly:
curl -sL "<url>" | python3 -c "
import sys
try:
    import fitz  # PyMuPDF
    doc = fitz.open(stream=sys.stdin.buffer.read(), filetype='pdf')
    print('\n'.join(page.get_text() for page in doc))
except ImportError:
    # Fallback: use pdftotext if available
    import subprocess
    result = subprocess.run(['pdftotext', '-', '-'], input=sys.stdin.buffer.read(), capture_output=True)
    print(result.stdout.decode())
"

From local file:

python3 -c "
import fitz
doc = fitz.open('$FILE_PATH')
for page in doc: print(page.get_text())
"

Fallback (no PDF tools): Use the summarize skill or web_fetch tool on the paper's URL. For arXiv papers, use the HTML version: https://arxiv.org/html/<id>.

2. Analyze with Structured Extraction

After obtaining the text, extract these categories:

## Paper: <title>
Authors: <authors>
Published: <date>
Source: <url>

### Core Contribution
<1-2 sentence summary of what's new>

### Key Techniques
- <technique 1>: <how it works, 2-3 sentences>
- <technique 2>: ...

### Agent Workflow Implications
- <how this applies to OpenClaw agent behavior>
- <specific workflow improvements suggested>

### Actionable Insights
1. <concrete thing we can implement>
2. <concrete thing we can implement>

### Skill Candidates
- <potential skill name>: <what it would do, trigger phrases>

### Limitations & Caveats
- <what doesn't apply or needs adaptation>

3. Generate Skill (if requested)

When asked to turn insights into a skill, use this scaffold:

---
name: <skill-name>
description: <one-line description derived from paper insight>
---

# <Skill Name>

Based on: <paper title> (<url>)

## When to Use
<trigger conditions>

## Technique
<extracted technique adapted for OpenClaw context>

## Workflow
<step-by-step procedure>

## Example
<concrete usage example>