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。
设计与 UI

deep-dive

Rigorous multi-agent deep-dive analysis for complex investigative tasks — auditing codebases, evaluating strategies or systems, validating designs, doing open-ended research. Deploys 4–6 specialist agents in parallel across distinct lanes, then synthesis, then adversarial red-team review, then optional patching — producing structured markdown research files plus a plain-English executive briefing with honest 1–10 confidence ratings. ALWAYS invoke when the user says any of "deep dive", "thorough audit", "rigorous analysis", "comprehensive review", "audit this codebase", "analyze the strategy", "evaluate this design", "review this thoroughly", or "research deep dive". Also invoke proactively for any open-ended investigative task involving a codebase, strategy, system design, or research question that warrants 30+ minutes of structured analysis — even when the user doesn't use these exact phrases.

nelsonwerd · 已有 AI 审核记录

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 nelsonwerd/idea-to-ship-skills --skill deep-dive

技能说明


name: deep-dive description: >- Rigorous multi-agent deep-dive analysis for complex investigative tasks — auditing codebases, evaluating strategies or systems, validating designs, doing open-ended research. Deploys 4–6 specialist agents in parallel across distinct lanes, then synthesis, then adversarial red-team review, then optional patching — producing structured markdown research files plus a plain-English executive briefing with honest 1–10 confidence ratings. ALWAYS invoke when the user says any of "deep dive", "thorough audit", "rigorous analysis", "comprehensive review", "audit this codebase", "analyze the strategy", "evaluate this design", "review this thoroughly", or "research deep dive". Also invoke proactively for any open-ended investigative task involving a codebase, strategy, system design, or research question that warrants 30+ minutes of structured analysis — even when the user doesn't use these exact phrases.

Deep-Dive Multi-Agent Analysis

This skill orchestrates rigorous multi-lane analysis for complex investigative tasks. It deploys specialist subagents in parallel, synthesizes their findings, runs adversarial red-team review, applies fixes, and delivers a structured evidence package plus a plain-English executive briefing.

When to use this

Strong triggers — invoke without asking:

  • "Do a deep dive on [X]"
  • "Thorough audit of [Y]"
  • "Rigorous analysis of [Z]"
  • "Comprehensive review"
  • "Audit this codebase"
  • "Evaluate this strategy / design"
  • "Research [open question] thoroughly"

Softer triggers — invoke if the task is investigative and non-trivial:

  • The user describes a codebase or system and asks for "thoughts" or "objective analysis"
  • The user has built something and asks whether it's correct/safe/sound
  • The user asks open-ended research questions that span multiple domains
  • The user is making a high-stakes decision and needs structured evidence

Do NOT use this for:

  • Single-file code review (use direct Read + analysis)
  • Simple factual questions (one WebSearch is sufficient)
  • Tasks the user has scoped tightly (e.g., "fix this bug" — just fix it)
  • Tasks under ~15 minutes of investigative work (overhead exceeds benefit)
  • Routing tie-breaker: deep-dive answers "is this correct / safe / viable / evidence-backed?" If the real question is "what should I build, or is this idea worth pursuing at all?" — a product-judgment call rather than an evidence audit — that's the ideate skill's job if you have it; if you don't, scope this run explicitly to viability evidence and say so. If the scope is already settled and the user just wants the build broken into sequenced steps, that's prompt-pack.

When in doubt, prefer to invoke — but right-size it. The skill scales from a quick 1–2-lane look to an exhaustive 6-lane investigation (see "Scale heuristics and run modes," which also covers the one-line cost heads-up before expensive runs).

The four scope variants

Identify which variant fits, then read the matching reference file:

VariantTrigger patternReference
Codebase audit"audit this codebase", "is this safe to ship", "review this build"references/codebase-audit.md
Strategy / system evaluation"evaluate this strategy", "does this approach actually work", "is this sound"references/strategy-evaluation.md
Research deep dive"research [open question]", "deep dive on [topic]"references/research-deep-dive.md
Design evaluation"evaluate this design", "is this the right approach", "review this plan"references/design-evaluation.md

If the user's request spans multiple variants (e.g., "audit this pricing-engine codebase") combine lanes from both reference files — the skill is composable. A codebase audit of a system with non-trivial decision logic uses codebase-audit lanes plus strategy/system-evaluation lanes for the logic itself.

If the scope is genuinely ambiguous, ask one or two clarifying questions before deploying agents. Don't ask a barrage. Examples:

  • "Pure research mode (no code changes) or are fixes in scope?"
  • "Focus on correctness, performance, or both?"

After the clarification, proceed.

Environment & fallbacks (run anywhere)

This skill is written for Claude Code, where parallel subagents and a few progress tools exist. The method is portable; only the orchestration mechanics degrade. The five stages — parallel/serial lanes → synthesis → follow-up verification → red-team → executive briefing — run in any capable agent (Claude Code, the Claude app, OpenAI Codex, or a generic agent). Before Phase 0, check what your runtime supports and substitute the fallback. Never tell the user "this only works in Claude Code" — adapt and run.

Claude-Code primitiveIf unavailable (Codex, a generic agent, or any runtime without it)
Parallel Agent calls in one message (Phase 1, Phase 3)Run the same lanes serially — one lane at a time, each writing its own markdown file, with the same prompts, deliverables, severity tiers, and confidence ratings. Same method, lanes, and deliverables — but one agent running lanes in sequence has less independence than separate agents cross-checking blind, so don't thin the lanes and don't let the final confidence read higher than that reduced independence supports. Longer wall-clock. This is the only acceptable substitute, and serial is correct here (see Pitfalls).
mcp__ccd_session__mark_chapter (Phase 0)Skip it. It's a progress signal, not part of the analysis.
TaskCreate task tracking (Phase 0)Skip it, or keep a short plain-text checklist in your reply. Progress sugar only.
WebSearch / WebFetch (Phase 1, follow-up)Rely on local artifacts (the repo, files, data the user provided). For any external/numerical claim you cannot verify locally, label it unverified — no web access inst
阅读完整来源 ↗