lastsearch: Skill for Claude Code

LastSearch — Evidence-Backed Research

Use this skill when the user needs researched, cited answers backed by real web sources — not LLM hallucinations.

When to Use

• User asks a factual question and wants verified, sourced answers
• User says "research this", "find out", "what does the evidence say", "look this up"
• User needs citations, confidence scores, or source verification
• User wants to know if something is true or needs fact-checking
• Any question where accuracy matters more than speed

Prerequisites

Install LastSearch MCP server:

{
  "mcpServers": {
    "lastsearch": {
      "command": "npx",
      "args": ["-y", "lastsearch"]
    }
  }
}

Or set LASTSEARCH_API_KEY=ls_xxx for full features (sessions, sharing, knowledge export).

Workflow

Step 1: Research the Question

Use answer to get a cited, evidence-backed answer:

answer({ query: "How do mRNA vaccines work?", depth: "fast" })

Use depth: "thorough" when:

• The topic is nuanced or controversial
• You need high confidence (thorough auto-retries with rephrased queries if confidence < 60%)
• The user explicitly asks for deep research

Step 2: Interpret the Response

The response contains:

• answer: The synthesized answer from real sources
• claims[]: Individual claims, each with source URLs, verification status, consensus level
• sources[]: Each source with URL, title, domain, quote, authority score
• confidence: 0-1 score computed from 7 real factors (NOT LLM self-assessed)
• contradictions[]: Conflicting claims found across sources
• trace[]: Pipeline timing (search, fetch, extract, verify, answer)

Step 3: Present to the User

When presenting results:

1. Lead with the answer
2. Cite sources inline using the URLs from claims[].sources
3. Mention confidence: "Confidence: 78% based on 5 sources"
4. If contradictions exist, surface them: "Note: sources disagree on X"
5. If confidence < 50%, caveat: "Limited evidence available — treat with caution"