A research is searching around something.

6-step structured research skill. Searches arXiv, reads 5 papers via markitdown, has Claude select the best 2, runs 4 targeted web searches (explanation, GitHub, survey, citations), then writes a 5-bullet summary. Use when a user asks for a "research summary", "deep dive", "what is X", "find papers on Y", "owl research", or any query that benefits from both academic papers and web context combined into a concise, citation-backed summary.

🦉 OWL — 6-Step Research Skill

A precise, structured research pipeline — not a bulk dump, but a curated read-and-select workflow that mirrors how an expert researcher actually works.

---

Pipeline

text

Step 1 · Search arXiv for the topic
Step 2 · Open 5 promising papers → PDF → Markdown via markitdown
Step 3 · Read title + abstract + intro of each
Step 4 · Claude selects the best 2 (explains why, drops the rest)
Step 5 · Search web: explanation · GitHub · survey · citations
Step 6 · Write a 5-bullet summary of what was learned

---

Install

bash

pip install requests markitdown
pip install curl_cffi beautifulsoup4 lxml fake-useragent
chmod +x scripts/owl.py
cp scripts/owl.py /usr/local/bin/owl

---

Usage

bash

owl "diffusion models"
owl "LoRA fine-tuning" --category cs.LG --output report.md
owl "protein structure prediction" --since year
owl "quantum error correction" --arxiv-n 8

For Windows user:

text

python  scripts/owl.py "AI agent RAG" --arxiv-n 8 --output report.md

---

CLI Flags

Flag	Default	Description
query	(required)	Research topic
--arxiv-n N	5	arXiv candidates to fetch (top N are used as pool)
--category CAT	all	arXiv category, e.g. cs.LG, q-bio.BM
--sort	relevance	relevance, lastUpdatedDate, submittedDate
--since RANGE	all time	Google recency: hour, day, week, month, year
--papers-dir DIR	/tmp/owl_papers	PDF + markdown cache directory
--output FILE	terminal only	Save Markdown summary to FILE
--no-stream	off	Wait for full Claude response
--serper-key KEY	env var	Override SERPER_API_KEY
--anthropic-key KEY	env var	Override ANTHROPIC_API_KEY

---

Step Details

Step 1 — arXiv Search

search_arxiv(query, n, category, sort_by) → list of paper dicts with id, title, authors, abstract, published, categories, url, pdf_url.

Step 2 — Open Papers (PDF → Markdown)

Take top 5 candidates. For each: download PDF to --papers-dir, convert to Markdown using markitdown (Python lib preferred, CLI fallback). Extract first 8 KB of text = roughly title + abstract + introduction. Falls back to abstract only if PDF download or conversion fails.

Step 3 — Read

Display truncated abstract+intro for each candidate. All 5 intro texts are passed to Claude in Step 4.

Step 4 — LLM Selects Best 2 papers

Claude returns {"selected": [i, j], "reason_1": "...", "reason_2": "...", "dropped": "..."}. The model used is printed in the step output and checklist. Selected papers then get their full markdown (up to 40 KB) loaded for Step 6.

Step 5 — 4 Web Searches

Runs 4 targeted Serper queries and fetches page content for the top 2 results of each:

Search type	Query pattern
explanation	{topic} explained
github	{topic} github implementation
survey	{topic} survey review paper
citations	{topic} highly cited papers results

Step 6 — 5-Bullet Summary

Builds a prompt with both full papers + all 4 web search result sets, then calls LLM api. Output structure:

1. Selected Papers (with links)
2. 5 dense bullets — each 3–5 sentences, specific facts, inline citations
3. Web Sources list

---

Output Format

markdown

# 🦉 OWL Research Summary: {topic}

## Selected Papers
[P1] Title — Authors — URL — PDF
[P2] Title — Authors — URL — PDF

## 5-Bullet Summary
• **What it is**: ...
• **How it works**: ...
• **Key results**: ...
• **Tools & code**: ...
• **Open questions**: ...

## Web Sources
...

---

Steps Checklist Verification

At the end of every run, owl prints a checklist confirming every step completed successfully. Verify all 6 items are marked [✓] before trusting the output.

text

────────────────────────────────────────────────────────────────────────
  ✅ RESEARCH CHECKLIST
────────────────────────────────────────────────────────────────────────
  [✓] Step 1  Search arXiv                  5 papers found
  [✓] Step 2  Open 5 papers via markitdown   5 PDFs converted
  [✓] Step 3  Read title + abstract + intro  5 papers scanned
  [✓] Step 4  Keep best 2                    2 selected  (2 with full text)
              → Paper title one…
              → Paper title two…
  [✓] Step 5  Web search (4 queries)         16 results fetched
              → explanation    Top result title…
              → github         Top result title…
              → survey         Top result title…
              → citations      Top result title…
  [✓] Step 6  5-bullet summary               written by Claude
              → /path/to/report.md
────────────────────────────────────────────────────────────────────────

What each line confirms

Step	What to verify
Step 1	At least 1 paper found — if 0, the arXiv query returned nothing; try rephrasing or removing --category
Step 2	PDFs converted — a ⚠ using abstract only warning here means markitdown failed for that paper; the run continues but that paper's selection is abstract-only
Step 3	Count matches Step 2 — should always be 5 (or --arxiv-n if overridden)
Step 4	Exactly 2 selected, model name visible — the model shown is whatever GET /v1/models returned as the best available; if it shows the fallback claude-sonnet-4-20250514 the models list API may have failed silently
Step 5	4 search types listed, each with a top result — a missing type means that Serper query failed silently
Step 6	"written by Claude" confirms the API call completed — if the summary is truncated, raise max_tokens in the source

Failure modes to watch for

• Step 2 all ⚠ — markitdown is not installed (pip install markitdown) or PDFs are blocked by arXiv rate limiting; wait and retry
• Step 4 shows 1 paper — Claude JSON parse failed; re-run or check OPENROUTER_API_KEY
• Step 5 shows 0 results for a type — Serper key exhausted or network issue; check SERPER_API_KEY
• Step 6 summary is very short — context window may be near limit with large PDFs; reduce --arxiv-n or use --no-stream

---

Notes

• PDFs are cached in --papers-dir — re-runs on the same paper are instant
• Step 4 uses a non-streaming JSON call; Steps 2 and 6 stream to terminal
• --arxiv-n can be raised to 8–10 for broader candidate pools before selection
• All 4 web search types always run — they cannot be individually disabled (by design)
• The 5-bullet format is enforced in the Claude prompt; free-form reports use the old pipeline