跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
搜索与研究

searcharvester-deep-research

Deep-research methodology with a ROLE-BASED team of parallel sub-agents. One `delegate_task` batch dispatches 2–3 researchers (each on a distinct sub-question), 1 critic (adversarial counter-search), and 1 fact-checker (numeric/date verification). The lead synthesises their findings into a cited report at ./report.md. Use for "research", "deep research", "report with sources", comparisons with citations, anything needing grounded multi-source evidence.

Searcharvester · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add vakovalskii/searcharvester --skill searcharvester-deep-research

技能说明


name: searcharvester-deep-research description: > Deep-research methodology with a ROLE-BASED team of parallel sub-agents. One delegate_task batch dispatches 2–3 researchers (each on a distinct sub-question), 1 critic (adversarial counter-search), and 1 fact-checker (numeric/date verification). The lead synthesises their findings into a cited report at ./report.md. Use for "research", "deep research", "report with sources", comparisons with citations, anything needing grounded multi-source evidence. version: 2.3.0 author: Searcharvester license: MIT metadata: hermes: tags: [research, deep-research, delegate_task, parallel, subagents, citations, adversarial] category: research related_skills: - searcharvester-search - searcharvester-extract - subagent-driven-development - writing-plans - plan

Searcharvester Deep Research (role-based)

Turn an open question into a cited markdown report grounded in real web sources. Instead of N homogeneous sub-agents, you dispatch a small TEAM with distinct roles — researchers gather, a critic tries to disprove, a fact-checker verifies numbers. Save to ./report.md.

Role

You are the lead researcher / editor. You do not personally run searcharvester-search or searcharvester-extract — the sub-agents do. Your job:

  1. Decompose the user query into 2–3 concrete sub-questions.
  2. Dispatch ONE delegate_task batch of 4–5 sub-agents with explicit roles (see Phase 2).
  3. Collect their findings, reconcile disagreements, write ./report.md.

When to use

  • "Research / deep research / analyse X"
  • "Compare A vs B with sources"
  • "What's publicly known about [person / topic / company]"
  • "Who holds the record for X" (factual questions benefit from the critic role — memory-based answers are often stale)
  • "What's the latest on X"

Do not use this skill for: coding tasks, or questions you genuinely cannot research online (private data, pure opinion).

Core principles

  • Evidence before prose. No claim in the final report without a source a sub-agent actually extracted.
  • Trust sources over memory. Even "well-known" facts get stale.
  • Adversarial verification. The critic's job is to be wrong on purpose — if they find contradictions, surface them.

Procedure

Phase 1 — Decompose (lead)

Write a plan to ./plan.md:

cat > ./plan.md << 'EOF'
## Intent
<one sentence on what the user wants>

## Sub-questions (2–3)
1. <concrete, factually researchable>
2. ...

## Critical facts to fact-check
- <specific number/date/name that needs verification>

## Out of scope
- <things we won't cover>
EOF

Phase 2 — Two-round pipeline (lead)

The team runs in TWO delegate_task rounds, not one big batch. The second round sees what the first round produced. Without this, the critic is just searching blind and often confirms whatever the model already "knows" from training.

Round 1 — Researchers only (parallel)

delegate_task(
    tasks=[
        {
            "goal": "Researcher: sub-question 1 — <sub-question text>",
            "context": RESEARCHER_TEMPLATE.replace("<SUBQ>", "<q1>").replace("<USER_QUERY>", user_query),
            "toolsets": ["terminal"],
        },
        {
            "goal": "Researcher: sub-question 2 — <sub-question text>",
            "context": RESEARCHER_TEMPLATE.replace("<SUBQ>", "<q2>").replace("<USER_QUERY>", user_query),
            "toolsets": ["terminal"],
        },
        # 2 to 3 researchers, one per sub-question
    ],
)

Wait for the results. Each researcher returned a ### Findings block with Claim/Quote/URL bullets, a ### Notes confidence line, and saved their extracts to ./extracts/*.md.

Between rounds — Prepare critic/fact-checker context (lead)

Collect the researchers' claims and facts. Write a short ### Researcher summary — one line per researcher, listing the top claim + URL — plus a ### Facts to verify list extracting every number, date, name, and record-count the researchers stated.

This block goes into the critic's and fact-checker's context so they can target SPECIFIC claims instead of searching blind.

Round 2 — Critic + Fact-checker (parallel, but with Round 1 in context)

delegate_task(
    tasks=[
        {
            "goal": "Critic: attack the researchers' conclusions",
            "context": (
                CRITIC_TEMPLATE
                  .replace("<USER_QUERY>", user_query)
                  .replace("<RESEARCHER_SUMMARY>", researcher_summary_block)
            ),
            "toolsets": ["terminal"],
        },
        {
            "goal": "Fact-checker: verify specific claims",
            "context": (
                FACT_CHECKER_TEMPLATE
                  .replace("<USER_QUERY>", user_query)
                  .replace("<FACTS>", facts_to_verify_block)
            ),
            "toolsets": ["terminal"],
        },
    ],
)

The critic now has concrete claims to challenge. The fact-checker has the exact numbers to verify. They can cat ./extracts/<id>.md to re-read the same extracts the researchers pulled (shared workspace).

Exactly two rounds. Do not fire a third. If a claim still looks shaky after round 2, surface it in the report's "Disagreements" section instead of starting another batch.


Template: RESEARCHER

ROLE: Researcher. You investigate ONE focused sub-question and return
a structured list of claims, each quoted from a real URL you read.

SUB-QUESTION:
<SUBQ>

PARENT USER QUERY (for context only):
"<USER_QUERY>"

TOOLS: You only have the `terminal` toolset. Call the searcharvester
scripts as shell commands — they are not registered as tools:

  # Search — returns JSON of URLs + snippets
  python3 /opt/data/skills/searcharvester-search/scripts/search.py \
    --query "<query>" --max-results 5

  # Extract — saves FULL markdown to ./extracts/<id>.md and returns a
  # pointer (id, url, total_chars, path, 800
阅读完整来源 ↗