Weeks 19-20: Auto Research Agent (2-Week Project)

This is your first multi-week project. You're going to build a research agent — an OpenClaw skill that takes a question, searches the web, reads pages, and writes a report with sources.

This combines everything from the last 6 weeks:

Agent loops (Week 13)
Tool use (Week 14)
Multi-step planning (Week 15)
Memory (Week 16)
OpenClaw skills (Weeks 17-18)

What You're Building

A skill that works like this:

You (via Telegram): Research: What are the pros and cons of nuclear energy?

Agent: I'll research that for you. Let me search for information...

[Agent searches the web, reads several pages, takes notes]

Agent: Here's what I found:

## Nuclear Energy: Pros and Cons

### Pros
- Low carbon emissions compared to fossil fuels (source: ...)
- High energy density (source: ...)
- Reliable baseload power (source: ...)

### Cons
- Radioactive waste storage (source: ...)
- High upfront construction costs (source: ...)
- Risk of accidents (source: ...)

### Sources
1. [URL 1]
2. [URL 2]
3. [URL 3]

Week 19: Build the Skill

Step 1: Understand the Research Pattern

The research agent follows this loop:

Question → User asks something
Search → Agent uses web_search to find relevant pages
Read → Agent uses web_fetch to read promising pages
Extract → Agent pulls out key facts and saves the source URL
Synthesize → Agent writes a report combining all facts
Cite → Agent includes source URLs

Step 2: Create the Skill on Your Server

SSH into your server:

ssh root@YOUR_IP_ADDRESS

Create the skill folder:

mkdir -p /opt/openclaw/workspace/skills/research_skill

Copy the template from this repo (research_skill/SKILL.md) or create it:

nano /opt/openclaw/workspace/skills/research_skill/SKILL.md

The skill file tells OpenClaw:

When to activate (user says "research:" or "look up:")
What tools to use (web_search, web_fetch)
How to format the output (sections, bullet points, sources)

Step 3: Test Basic Searches

Try simple queries first:

openclaw chat "Research: What is the tallest building in the world?"

Watch the output. Does it search? Does it cite sources?

Step 4: Iterate

The first version won't be perfect. Common issues:

Agent doesn't search enough sources → Update SKILL.md to say "search at least 3 sources"
Sources aren't cited → Add explicit instructions about including URLs
Report is too short → Ask for "at least 200 words with specific facts"

Week 20: Improve and Polish

Make It Better

Better search queries: Add instructions for the agent to rephrase the question as a good search query
Multiple perspectives: Tell the agent to search for different viewpoints
Fact checking: Add a step where the agent verifies key claims across sources
Better formatting: Use headers, bullet points, and bold for key terms

Test with Harder Questions

Try the questions in example_queries.md. These are designed to test different aspects:

Questions that need multiple searches
Questions with controversial answers
Questions that need recent information

Share It

Send research questions to your agent via Telegram. Share the bot with a friend and have them try it.

Tasks

Week 19:

Create the research skill on your OpenClaw server
Test with 3 simple research questions
Iterate on the SKILL.md at least twice based on results

Week 20:

Test with the questions in example_queries.md
Improve the skill based on what you learned
Have someone else try it and get feedback
Fill in reflection.md

Deliverables

Your research_skill/SKILL.md (copy it to this repo folder)
Screenshots of 3 research runs (including one from example_queries.md)
reflection.md filled in

Stuck?

"I'm building a research skill for OpenClaw. The agent [searches but doesn't synthesize / doesn't cite sources / gives shallow answers]. Here's my SKILL.md: [paste]. How can I improve it?"

Remember: most problems with OpenClaw skills are solved by making the SKILL.md instructions clearer and more specific.

README.md