원클릭으로 Manus에서 모든 스킬 실행

deep-research

스타8

포크1

업데이트2026년 6월 15일 02:31

Deep research harness — fan-out web searches, fetch sources, adversarially verify claims, synthesize a cited report. When the user wants a deep, multi-source, fact-checked research report on any topic. BEFORE invoking, check if the question is specific enough to research directly — if underspecified (e.g., "what car to buy" without budget/use-case/region), ask 2-3 clarifying questions to narrow scope. Then pass the refined question as args, weaving the answers in.

설치

Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.

Manus에서 실행

출처

ek33450505

ek33450505/claude-agent-team

GitHub 저장소 열기 Creator 저장소 보기

다운로드

Manus에서 실행

Deep Research Skill

A multi-agent research harness: it decomposes a question into search angles, fans out parallel web searches, fetches and extracts falsifiable claims from sources, adversarially verifies every claim with a 3-vote skeptic panel, and synthesizes a cited report.

The workflow lives in deep-research.workflow.js next to this file. It is the single source of truth for the harness — edit it there, then (if you keep a repo copy) run bash install.sh to redeploy.

Usage

/deep-research <your research question>

How to run it

Scope-check first. If the question is underspecified (missing budget, region, use-case, timeframe, etc.), ask the user 2-3 clarifying questions before running. Then fold the answers into a single refined question string.
Invoke the workflow via the Workflow tool, pointing scriptPath at the workflow file in this skill's directory and passing the refined question as args:
```
Workflow({
  scriptPath: "~/.claude/skills/deep-research/deep-research.workflow.js",   // expand ~ to $HOME
  args: "<the refined research question>"
})
```
Pass the question as a plain string in args (not JSON-encoded). The workflow reads it via the global args. This skill's instruction to call Workflow is itself the explicit opt-in for multi-agent orchestration.
Relay the result. The workflow returns { summary, findings, refuted, unverified, sources, stats }. Present the summary and findings with their sources; surface unverified claims clearly tagged as unverified (verification couldn't render a verdict — NOT refuted), and keep refuted claims in a transparency section.

Output contract (important)

The harness reports three claim outcomes, never two:

confirmed — a quorum of valid votes, fewer than 2 refuting.
refuted — a quorum of valid votes with ≥2 actually voting to refute.
unverified — too few valid verdicts to adjudicate (agent/tooling failure or abstentions). These are returned with their sources and an explicit "unverified" tag. Never present an unverified claim as confirmed, and never report it as refuted. A run where verification couldn't execute returns the extracted claims marked unverified — it does not collapse to an empty findings list with a "refuted" headline.

stats.fallbacksFired / stats.fallbackRecovered show how often a low-fan-out subagent (scope/ search/fetch/synthesize) failed to emit structured output and the free-text→structure fallback recovered it. A high unverified count with confirmed: 0 means verification couldn't run (usually server load) — say so plainly when relaying results; do not dress it up as a finding.

Design notes

Cost scales with the question, dialed back for load (2026-06-08). Default fan-out: 5 angles, ≤10 fetched sources, top-12 claims × 3 verifier votes (~50–65 subagent calls). Verify is the heaviest phase, so it runs in sequential batches (VERIFY_BATCH claims each, direct call) to pace requests under the server's transient rate limiter — each batch fully drains before the next.
Resilience, by phase:
- Verify uses a direct structured call with no fallback amplification. A live post-mortem (run wf_2b84c096) showed the free-text→structure fallback only multiplied load under rate-limiting (75 votes → 225 agents, 0 recovered). A failed/rate-limited vote → abstain → the claim is unverified (never refuted) — honest degradation, not a fabricated verdict.
- Scope / Search / Fetch / Synthesize (low fan-out) keep structuredWithFallback(): if a subagent does its web work then ends without StructuredOutput, the work is re-run as a free-text pass and a structure-only agent renders it — so one hiccup doesn't drop a good source.
Honest failure is the contract. When the service is busy the harness returns unverified claims and says so; it does not fabricate confirmations or refutations. See the two 2026-06-08 post-mortem notes in the workflow header.

이 저장소의 다른 Skills

같은 저장소

cast-conventions

ek33450505/claude-agent-team

Shared CAST conventions for all agents. Loaded automatically via agent frontmatter.

2026-07-188

orchestrate

ek33450505/claude-agent-team

Execute a CAST plan by reading the Agent Dispatch Manifest and dispatching agents in dependency order — parallel batches simultaneously, sequential batches one at a time. Pass a plan file path, 'next' for the most recent plan, or 'resume' to continue from a checkpoint.

2026-07-188

agent-registry

ek33450505/claude-agent-team

Full CAST agent registry with model assignments. Load when building an Agent Dispatch Manifest, orchestrating multi-agent workflows, or when you need to verify which model a specific agent runs on.

2026-07-118

cast-audit

ek33450505/claude-agent-team

Comprehensive CAST codebase audit — bugs, security, performance, test coverage. Dispatches 4 parallel researchers to scan the codebase and collates findings into a dated report. Manual invocation only; monthly scheduling is a design intention (helper not yet shipped).

2026-07-068

briefing-writer

ek33450505/claude-agent-team

Assemble morning briefing sections into a structured markdown file. Use when writing the final briefing output after all data sources have been gathered.

2026-06-158

careful-mode

ek33450505/claude-agent-team

Activate careful mode — require explicit confirmation before any Write, Edit, or Bash operations. Use when working on production code or sensitive systems.

2026-06-158

name	deep-research
description	Deep research harness — fan-out web searches, fetch sources, adversarially verify claims, synthesize a cited report. When the user wants a deep, multi-source, fact-checked research report on any topic. BEFORE invoking, check if the question is specific enough to research directly — if underspecified (e.g., "what car to buy" without budget/use-case/region), ask 2-3 clarifying questions to narrow scope. Then pass the refined question as args, weaving the answers in.
user-invocable	true
allowed-tools	["Workflow","Read"]

Deep Research Skill

Usage

/deep-research <your research question>

How to run it

Scope-check first. If the question is underspecified (missing budget, region, use-case, timeframe, etc.), ask the user 2-3 clarifying questions before running. Then fold the answers into a single refined question string.
Invoke the workflow via the Workflow tool, pointing scriptPath at the workflow file in this skill's directory and passing the refined question as args:
```
Workflow({
  scriptPath: "~/.claude/skills/deep-research/deep-research.workflow.js",   // expand ~ to $HOME
  args: "<the refined research question>"
})
```
Pass the question as a plain string in args (not JSON-encoded). The workflow reads it via the global args. This skill's instruction to call Workflow is itself the explicit opt-in for multi-agent orchestration.
Relay the result. The workflow returns { summary, findings, refuted, unverified, sources, stats }. Present the summary and findings with their sources; surface unverified claims clearly tagged as unverified (verification couldn't render a verdict — NOT refuted), and keep refuted claims in a transparency section.

Output contract (important)

The harness reports three claim outcomes, never two:

confirmed — a quorum of valid votes, fewer than 2 refuting.
refuted — a quorum of valid votes with ≥2 actually voting to refute.
unverified — too few valid verdicts to adjudicate (agent/tooling failure or abstentions). These are returned with their sources and an explicit "unverified" tag. Never present an unverified claim as confirmed, and never report it as refuted. A run where verification couldn't execute returns the extracted claims marked unverified — it does not collapse to an empty findings list with a "refuted" headline.

Design notes

Cost scales with the question, dialed back for load (2026-06-08). Default fan-out: 5 angles, ≤10 fetched sources, top-12 claims × 3 verifier votes (~50–65 subagent calls). Verify is the heaviest phase, so it runs in sequential batches (VERIFY_BATCH claims each, direct call) to pace requests under the server's transient rate limiter — each batch fully drains before the next.
Resilience, by phase:
- Verify uses a direct structured call with no fallback amplification. A live post-mortem (run wf_2b84c096) showed the free-text→structure fallback only multiplied load under rate-limiting (75 votes → 225 agents, 0 recovered). A failed/rate-limited vote → abstain → the claim is unverified (never refuted) — honest degradation, not a fabricated verdict.
- Scope / Search / Fetch / Synthesize (low fan-out) keep structuredWithFallback(): if a subagent does its web work then ends without StructuredOutput, the work is re-run as a free-text pass and a structure-only agent renders it — so one hiccup doesn't drop a good source.
Honest failure is the contract. When the service is busy the harness returns unverified claims and says so; it does not fabricate confirmations or refutations. See the two 2026-06-08 post-mortem notes in the workflow header.