SKILL.md — Web Research

Search the web and synthesize findings into a structured report.

---

Purpose

Research a topic using web search, fetch relevant sources, and produce a synthesized summary with citations.

---

Triggers

• "research {topic}"
• "find out about {topic}"
• "look up {topic}"
• "search for {topic}"
• When user asks questions requiring current information

---

Trust Required

L1 (Observe) — This skill is read-only.

---

Inputs

Name	Type	Required	Description
topic	string	yes	What to research
depth	string	no	"quick" (3 sources) or "deep" (10 sources). Default: quick
focus	string	no	Specific angle or question to answer

---

Outputs

Name	Type	Description
summary	markdown	Synthesized findings (200-500 words)
key_points	list	3-5 bullet points with main takeaways
sources	list	URLs with titles, used in research
confidence	string	"high", "medium", "low" based on source quality

---

Implementation

Step 1: Generate Search Queries

Based on the topic, generate 3-5 search queries:

• Main topic query
• Specific aspect queries
• Recent/news query (add "2026" or "latest")

Example for "AI agent security":
1. "AI agent security best practices"
2. "prompt injection defense 2026"
3. "LLM agent vulnerabilities OWASP"
4. "AI agent sandboxing techniques"

Step 2: Execute Searches

Use web_search for each query:

results = []
for query in queries:
    search_results = web_search(query, max_results=3)
    results.extend(search_results)

Step 3: Deduplicate and Rank

Remove duplicate URLs, prioritize:

1. Official documentation
2. Peer-reviewed or authoritative sources
3. Recent publications (prefer 2025-2026)
4. Avoid: forums, social media, SEO spam

Step 4: Fetch Top Sources

Use web_fetch on top 3-10 sources (based on depth):

contents = []
for source in top_sources:
    content = web_fetch(source.url)
    contents.append({
        "url": source.url,
        "title": content.title,
        "text": content.content[:5000]  # Truncate if huge
    })

Step 5: Synthesize

Combine information from all sources:

• Identify common themes
• Note contradictions
• Extract key facts
• Form coherent narrative

Step 6: Format Output

## Research: {topic}

### Summary
[200-500 word synthesis]

### Key Points
- Point 1
- Point 2
- Point 3

### Sources
1. [Title](url) — Brief note on what this contributed
2. [Title](url) — Brief note
3. [Title](url) — Brief note

### Confidence: {high/medium/low}
[Brief note on source quality and coverage]

---

Example

Input:

research "OWASP LLM Top 10 2025" depth:quick

Output:

## Research: OWASP LLM Top 10 2025

### Summary
The OWASP LLM Top 10 2025 identifies the most critical security 
risks for applications using large language models. Prompt injection 
remains the #1 threat, now categorized as both direct (user-supplied) 
and indirect (embedded in external content). New entries include 
"Agentic Risks" reflecting the rise of autonomous AI agents with 
tool access...

### Key Points
- Prompt injection is #1, with indirect injection most dangerous for agents
- "Agentic Risks" is new, covering tool misuse and privilege escalation
- Data poisoning moved up to #3 due to RAG vulnerabilities
- Recommendations emphasize defense-in-depth and human oversight

### Sources
1. [OWASP LLM Top 10 2025](https://owasp.org/...) — Primary source
2. [LLM Security Landscape 2025](https://...) — Industry analysis
3. [Agentic AI Security Guide](https://...) — Deep dive on agent risks

### Confidence: High
Multiple authoritative sources with consistent findings.

---

Error Handling

Error	Response
No search results	Try broader query, report if still empty
All fetches fail	Return search snippets only, note limitation
Contradictory sources	Present both views, note disagreement
Topic too broad	Ask user to narrow focus

---

Notes

• Always cite sources
• Prefer recent information (2025-2026)
• If topic is time-sensitive, note when sources were published
• For controversial topics, present multiple perspectives
• Never present speculation as fact