{"id":"geek-skills-deep-research","name":"deep-research","summary":"このスキルは、エビデンスに基づく研究メモ、文献レビュー、市場・政策・技術的状況、または引用、トレードオフ、明確な結論を含む多情報源の意思決定ブリーフを求める際に活用してください。","body":"# Deep Research V8.1\n\nThis skill is for **evidence-rich research outputs**, not for every question that happens to mention “analysis”.\n\nThe V8 shift is simple:\n- **Single-agent first.** Start with one lead agent and only fan out when parallel work will clearly help.\n- **Thin harness, fat skill.** Put reusable judgment and workflow here; keep deterministic checks in scripts.\n- **Context organization over prompt stuffing.** Load the minimum active context bundle, then pull in references only when needed.\n- **Eval and observability built in.** A good report is not enough; the run must also be diagnosable and improvable.\n\n## What this skill should produce\n\nChoose the lightest artifact that satisfies the task.\n\n| Output type | Use when | Typical length | Required artifacts |\n|---|---|---:|---|\n| **Brief memo** | user wants a concise answer with evidence | 800-1800 words | `research-plan.md`, `registry.md`, `draft.md`, `run-summary.json` |\n| **Full report** | user asks for comprehensive analysis / literature review / decision document | 2500-6000 words | all core artifacts + `evaluation.md` |\n| **Delta update** | user says “continue”, “second round”, “what changed”, “deepen round 2” | 600-1800 words | prior round handoff (`references/handoff-format.md`) + new notes + delta draft |\n\nIf the user did **not** ask for a long report, default to **Brief memo**.\n\n## When NOT to use this skill\n\nDo **not** activate for:\n- quick fact lookups or simple definitions\n- summarizing a single provided article/PDF/page\n- short comparisons the model can answer directly from 1-2 sources\n- brainstorming without evidence requirements\n- tasks where the user explicitly wants a short answer, not a report\n\nIf in doubt, ask yourself: **Does this task need a reusable evidence artifact and multi-source synthesis?** If not, do something simpler.\n\n## Org-policy boundary\n\nThis skill does **not** replace system policies, enterprise guardrails, or repo-level instructions.\nPut these outside the skill:\n- data handling / PII / compliance rules\n- approval requirements for external access or irreversible actions\n- org-wide style and review policy\n- environment-specific permissions\n\nKeep those in system prompts, AGENTS/CLAUDE/OpenAI config, or the harness. This skill owns the **workflow**, not the company’s permanent red lines.\n\n## Active context bundle\n\nAt activation time, keep the active bundle small.\n\n**Always load first**\n1. This `SKILL.md`\n2. `references/methodology.md`\n3. `references/report-assembly.md`\n4. `references/research-notes-format.md`\n\n**Load on demand**\n- `references/subagent-prompt.md` only if you actually dispatch subagents\n- `references/handoff-format.md` only when a delta update continues a prior round\n- `references/evaluator-prompt.md` only if you run the evaluator\n- `references/quality-gates.md` before finalization\n- `references/observability.md` when emitting metrics or diagnosing regressions\n- `references/tension-discovery.md` only for contested / decision-heavy topics\n- `references/landscape-scan.md` only when literature or ecosystem mapping matters\n\n**After compaction or context reset**\nReload only:\n- `research-plan.md`\n- active task notes\n- `registry.md`\n- unresolved issues list\n- the one reference file for the current phase\n\nDo **not** reload the whole skill tree unless the run drifted badly.\n\n## Workflow\n\n### P0 — Scope, route, and choose the lightest mode\n\nCreate `workspace/research-plan.md` with:\n- research question\n- intended audience\n- freshness requirement\n- geography / market / jurisdiction\n- output type (brief / full / delta)\n- stakes: low / medium / high\n- why this skill is justified\n\nThen choose the orchestration mode:\n\n| Mode | Default choice |\n|---|---|\n| **Single-agent** | default for most tasks |\n| **Lead + subagents** | only when there are 3+ separable research threads or obvious parallel value |\n| **Delta update** | when continuing prior research |\n\n**Do not fan out just because subagents exist.**\n\n### P0.5 — Optional modules (not mandatory by default)\n\nUse optional modules only when they earn their keep:\n- **Tension discovery** (`references/tension-discovery.md`): use for contested, hype-heavy, or decision topics where mainstream framing may be wrong.\n- **Landscape scan** (`references/landscape-scan.md`): use when the domain is unfamiliar, broad, or literature-heavy. For non-academic topics, this can be an ecosystem/standards/vendor scan rather than arXiv.\n- **Reverse search**: use when costs, failure modes, counter-evidence, or operational constraints are missing.\n\n### P1 — Plan the evidence work\n\nBreak the task into 1-5 research threads. Each thread needs:\n- one crisp objective\n- starting queries\n- what “done” looks like\n- what evidence would change the conclusion\n\nIf using subagents, each subagent gets **one** focused thread. Avoid overlapping ownership.\n\n### P2 — Investigate, extract, and write notes\n\nFollow `references/research-notes-format.md`.\n\nRules:\n- search broadly first, then chase named entities, standards, datasets, products, trials, laws, or papers\n- fetch and read the best supporting sources for the highest-value claims\n- write notes that separate **facts**, **analysis**, **gaps**, and **unresolved conflicts**\n- capture support snippets/paraphrases for the top claims so later verification is easier\n\nThe lead agent should work from notes **by default**, but may inspect raw/fetched sources again when:\n- two sources materially conflict\n- a claim is high-stakes or decision-critical\n- a note looks suspiciously weak or over-compressed\n\n### P3 — Build registry and verify evidence\n\nCreate `workspace/registry.md` from approved sources only.\n\nUse `scripts/source_evaluator.py` as a **helper**, not an oracle.\nAuthority scores are heuristics. Final acceptance depends on claim fit, evidence type, and whether the source can actually bear the weight of the claim.\n\nUse `scripts/verify_citations.py` before finalization.\n\nEvidence rules:\n- core claims should lean on the strongest available evidence for that claim type\n- anecdotes illustrate; they do not anchor the conclusion\n- conflicting evidence must be surfaced, not silently averaged away\n- if the topic is high-stakes, spot-check raw support for top claims before shipping\n\n### P4 — Synthesize the output\n\nFollow `references/report-assembly.md`.\nFor full reports, `assets/report_template.md` is an optional skeleton; adapt it rather than inventing a new structure.\n\nAlways include:\n- clear answer to the user’s question\n- explicit limitations / trade-offs\n- separation of source-backed findings vs your own synthesis\n- uncertainty calibrated to evidence quality\n\nOnly include a dedicated **Decision Framework** when the user is choosing between options.\nOnly require a **contrarian** section when the topic actually has a mainstream narrative worth challenging. Otherwise produce a **non-obvious insight** instead of forcing fake contrarianism.\n\n### P5 — Evaluate and gate\n\nFor full reports and medium/high-stakes briefs, run the evaluator using `references/evaluator-prompt.md`.\n\nBefore finalization, check `references/quality-gates.md`:\n- routing correctness\n- process completeness\n- grounding / citation integrity\n- output quality\n- efficiency and operational health\n\n### P6 — Finalize, summarize, and learn\n\nEmit:\n- final `draft.md`\n- `evaluation.md` if run\n- `run-summary.json` via `scripts/emit_run_summary.py`\n\nIn the run summary, record what actually helped: single-agent, subagents, tension discovery, landscape scan, reverse search, evaluator, or manual spot-checks.\nThis is what makes the skill improve over time.\n\n“Finalize” means deliver research artifacts to the user. Do not publish to an\nexternal site, send messages, request credentials, or make irreversible changes\nunless a separate explicit user request and the host policy authorize it.\n\n## Deterministic helpers\n\nUse scripts for the parts that should be boring and repeatable:\n- `scripts/source_evaluator.py` — baseline source scoring / diversity checks\n- `scripts/verify_citations.py` — fail-closed citation integrity and source-pool checks\n- `scripts/emit_run_summary.py` — structured observability output for the run\n\nFor reproducible recency scoring, pass an explicit evidence cutoff:\n`scripts/source_evaluator.py sources.json --as-of YYYY-MM-DD`. Record the same\ncutoff in the research plan and run summary.\n\nIf a deterministic check fails, fix the artifact first. Do not argue with the script unless you have a concrete reason.\n\n## Evaluation and observability\n\nThis skill is only “good” if it performs well on:\n1. **Routing** — does it trigger when it should, and stay out of the way when it should not?\n2. **Process** — did it create the right artifacts and evidence trail?\n3. **Outcome** — is the final brief/report genuinely useful and grounded?\n4. **Efficiency** — did it get there with acceptable tool/time/token cost?\n5. **Safety / governance** — did it respect policy boundaries and handle uncertainty honestly?\n\nSee:\n- `evals/routing-evals.json`\n- `evals/runbook.md` — how to run the routing and mode checks\n- `references/quality-gates.md`\n- `references/observability.md`\n\n## Degraded mode\n\nIf subagents, shell, or a writable workspace are unavailable, keep the workflow but shrink the surface area:\n- one lead agent only\n- inline notes instead of files if needed\n- fewer searches, but still enough to support the conclusion\n- lightweight evaluator or self-check if full evaluation is impossible\n- still keep limitations, uncertainty, and citation integrity\n\n## Stop conditions\n\nStop and ask for help only when the blocker is real and specific, for example:\n- no credible sources exist for a critical claim\n- the user’s requested scope conflicts with available evidence\n- policy or access restrictions block the required research\n\nOtherwise, continue with the best justified artifact and say where the confidence drops.","author":"@staruhub","ownerProfile":null,"authorContacts":null,"sourceUrl":"https://github.com/staruhub/ClaudeSkills/tree/main/skills/Geek-skills-deep-research","license":"MIT","category":"document","lang":"en","tokens":2162,"stars":0,"calls30d":1,"claimed":false,"visibility":"public","origin":"crawler","version":"0.1.0","createdAt":"2026-08-22","updatedAt":"2026-08-22","files":[{"path":"assets/report_template.md","size":1006,"sha256":"0f5340b19131f74a3212bc14db55256f7dd91476d48a156b47ec9025bf1f1ab6"},{"path":"evals/routing-evals.json","size":3859,"sha256":"9ab89286d525aa07d9883b596d701cf324e761cb5f1be5f1aa3bea3e5a164342"},{"path":"evals/runbook.md","size":967,"sha256":"3ce10c7fc6c8e5c2153fa2783d264ad67cbd0b07de17f7c3205fa7fd4251d204"},{"path":"references/evaluator-prompt.md","size":4090,"sha256":"d46c73603c124f62c84e2cf7303b7ddbbd216fb9682294b0f003202ff8caa21a"},{"path":"references/handoff-format.md","size":4027,"sha256":"6b1595832b9217f42a6a9d699ff3247692b5fb55f741f5bc229fc593e247c566"},{"path":"references/landscape-scan.md","size":1194,"sha256":"165ea93f64c21b949fccdcedf25ab206a2906ea471fd938c6e11d1749f6a33c5"},{"path":"references/methodology.md","size":5852,"sha256":"4075b47a6fec26ee075ce2cc197ace20efa7be2918202422d265ba1280f65e1c"},{"path":"references/observability.md","size":3051,"sha256":"9f25c2d95e28e6d1b699ca2feffad2f0f1c507099168b61cdfd85a03e5fe039c"},{"path":"references/quality-gates.md","size":5593,"sha256":"c8b87b4e7baf39a9626fc91869c68cb93bac86a4dc6641deed928e60ef38ab4e"},{"path":"references/report-assembly.md","size":3435,"sha256":"20656c9a10846c74574b310e05351b1246ef2b674139a939f7242a6dc7d688e2"},{"path":"references/research-notes-format.md","size":2203,"sha256":"cc2e38cbd5ca8d442c52ac14e78782f6370c8da77617c304a3e67a9c467ea032"},{"path":"references/subagent-prompt.md","size":1787,"sha256":"3b152fb5e046b63e9f3e65dd735e2dbff756b4d1f3ed89117a376abe700b8169"},{"path":"references/tension-discovery.md","size":1235,"sha256":"75bf29d3b6c074ede1d17da6e8865a1d3201145103f7e91b0160f005d8922b81"},{"path":"scripts/emit_run_summary.py","size":5129,"sha256":"4fed8d369957c48f87cc987fdc8b2bfa17375862a2c94899b3b59051220f52a4"},{"path":"scripts/source_evaluator.py","size":12234,"sha256":"19d0f905b8a5f415bd07ee98bae69edc7cb91d424306d9fac45ec4264e363967"},{"path":"scripts/verify_citations.py","size":11006,"sha256":"aeed81a6757595e3006c2747a8af4e43eca5130fe1f18d046def44273a2e45f2"}],"requires":{"mcp":[],"tools":[]},"safety":{"flags":[],"scannedAt":"2026-08-22","hasScripts":true,"networkEndpoints":[]}}