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>

Write the skill to ~/.openclaw/workspace/skills/<skill-name>/SKILL.md.

4. Save Insights

Store extracted insights for future reference:

# Append to the research log
cat >> ~/.openclaw/workspace/memory/research-insights.md << 'EOF'

## <Paper Title> (<date analyzed>)
Source: <url>
Key insight: <1-liner>
Applied to: <skill name or workflow>
EOF

Output Format

Always structure output as:

📄 Paper: <title>

🔬 Core Contribution
<summary>

⚡ Actionable for OpenClaw
1. <action item with expected impact>
2. <action item>

🛠️ Skill Potential: <Yes/No>
<if yes, skill name + description>

📝 Next Step
<specific command or action to take>

Examples

Input: "Analyze the ParaThinker paper on parallel reasoning"

Output: Analysis showing parallel chain-of-thought technique → suggests spawning multiple reasoning paths as subagents → generates a parallel-reasoning skill scaffold.

Input: "What can we learn from this PDF about code review?"

Output: Extracts review heuristics → maps to CodeFlow's review phase → suggests phase-aware prompt improvements.

Dependencies

• python3 with PyMuPDF (pip3 install pymupdf) — preferred
• Or pdftotext (from poppler) — fallback
• Or summarize skill — for URL-based papers
• Or web_fetch tool — for HTML papers (arXiv)

Install PDF support:

pip3 install pymupdf