English · 简体中文
A "deep research" skill for Claude Code: works like a senior consultant — understand the question → produce a research plan → verify iteratively across multiple sources (with built-in dissent and cross-checking) → deliver a report with a judgment, evidence, and traceable sources.
A plain Q&A easily hands you an answer that sounds right. Deep research is different: it forms hypotheses, dispatches several independent researchers to verify in parallel, actively hunts for counter-evidence, checks recency, and ends with a judgment that carries a confidence level and "under what conditions this conclusion would be wrong." Good for selection, due diligence, comparison, and investment/purchase decisions — anything you can't afford to wing.
- Research plan — restate your question + a tentative take + 3-5 angles + expected depth, shown to you first.
- Clear-direction questions with nothing for you to steer (diagnostic / verification / status, e.g. "why does X do Y") → no plan, no prompt, it just starts.
- When direction is open (survey / selection / scope is yours to set), it shows the plan and auto-starts 60 seconds later; during that window you can change the plan, or reply "start now" to go immediately.
- Put "立即开始 / start now" in the command from the outset to skip the wait entirely.
- Autonomous research — dispatches sub-agents to search, extract, cross-verify and find dissent in parallel; decides whether to run another round based on saturation. No interruptions once started.
- Deliver report — consultant-grade analysis (central thesis, judgment, actionable next steps), not a dump of search results.
# 1) Drop it into the Claude Code skills directory
cp -r deep-research ~/.claude/skills/deep-research
# 2) (Optional) install the /deep-research command alias: put the command file into ~/.claude/commands/Then in Claude Code just say "deep research <topic>" or /deep-research \<topic\>.
| Tier | Time | researchers | Words | For |
|---|---|---|---|---|
| fast | 5-10 min | 0 (main agent) | 1000-2500 | single-point deepening / decision memo |
| standard (default) | 15-30 min | ≥2 | 2500-5000 | medium research / selection |
| deep | 40-60 min | ≥4 | 5000-12000 | systematic research / major decisions |
If any hard constraint isn't met, the report honestly flags degraded up front — it never pretends to have done the full job.
- No fabrication — every number has a traceable source; if unsure, say so.
- Find the other side — actively search counter-views and failure cases; one-sided = not done.
- Mark the gaps — "I don't know X, so confidence in Y is only Z."
- Give a judgment — central thesis, confidence, actionable next step.
- Be falsifiable — every core conclusion ships with "under what conditions it's wrong."
After changing the skill itself, run the built-in harness:
bash hooks/verify.sh all # fixture self-tests + cross-file consistency + skill integrity| File | Purpose |
|---|---|
SKILL.md |
workflow + hard constraints + golden rules (single source of truth) |
PLAYBOOK.md |
per-phase execution detail, researcher prompts, saturation criteria |
REPORT_TEMPLATES.md |
survey / compare / decide / howto report skeletons |
RUBRIC.md |
Devil's Advocate + self-scoring |
WRITING_STYLE.md |
consultant-grade writing standard |
experts/ |
per-domain quality-dimension templates |
hooks/ |
output-validation scripts + self-check harness |
An engineering combination of several published research methods (understand the spirit, apply as needed, no name-dropping): ReAct, Reflexion, Self-Ask, CRAG, Step-Back, the Toulmin argument model, plus targeted re-checking for "later sections hallucinate more."