deep-research

name	deep-research
description	Multi-round research with explicit methodology, evidence tracking, and citation-tagged synthesis. Trigger on 'deep dive', 'research report', 'literature review', 'investigate X across sources', 'multi-round investigation'. Distinct from the `summarize` skill, which is a single-pass condensation; this skill maintains a state file across iterations, tracks coverage, and produces a long-form report with per-claim citations. Three execution stages: plan (scope into sub-questions), iterate (record evidence per round), compile (synthesize report). The skill itself does not fetch the web — it tells the host agent which fetches to perform via OpenSquilla's existing web tools, and records what comes back.
description_zh	具有明确方法论、证据追踪和带引用标注综合的多轮研究。触发词如'深入研究'、'研究报告'、'文献综述'、'跨来源调查X'、'多轮调查'。区别于单轮浓缩的 summarize 技能：本技能跨迭代维护状态文件、追踪覆盖度，并生成带逐条引用的长篇报告，分为规划、迭代、汇编三个阶段。技能本身不抓取网页，而是通过OpenSquilla现有的网页工具告知主机代理执行抓取并记录结果。
homepage
provenance	{"origin":"clawhub-mit0","license":"MIT-0","upstream_url":"https://clawhub.ai/in-depth-research","maintained_by":"OpenSquilla"}
metadata	{"platform":{"emoji":"🔬"}}

deep-research

Investigate a question by walking it through three explicit stages with a persisted state file. Use this when a single-pass summarize would lose too much, or when the user asks for a "research report" / "literature review". The host agent does the web fetching; this skill structures the work and keeps a paper trail.

Decide if this is the right tool

Need	Use
One-line summary of an article	summarize
Multi-round investigation with citations	this skill
Quick lookup, single source	direct web search
Continuous monitoring of a topic	a digest/cron skill

Stages

Scope → Plan → Iterate (×N) → Compile → Deliver

State persists in a single JSON file you pass between stages. The file is the contract; if you can describe the file, you can resume the research at any point.

---

Stage 1: Plan

python {baseDir}/scripts/plan.py \
    --question "How did Manus differentiate from competing AI agents in 2025?" \
    --depth thorough \
    --out plan.json

--depth choices:

• overview — 3-5 sub-questions, target 1 source per sub-question
• thorough — 6-10 sub-questions, target 2-3 sources per sub-question
• exhaustive — 12-20 sub-questions, target 5+ sources per sub-question

The plan is a pydantic model serialized to JSON; see references/methodology.md for the schema and the system-review approach the depth choices implement.

---

Stage 2: Iterate

Each round: read the plan, decide which sub-questions need attention, print the fetch list for the host agent to execute, and (after the agent returns results) record evidence back into the plan.

# Show the host what to fetch this round
python {baseDir}/scripts/iterate.py --plan plan.json --round 1 --print-fetches

# After the host fetches, record results back
python {baseDir}/scripts/iterate.py --plan plan.json --round 1 \
    --record evidence_round_1.json