메뉴
Use when a task requires systematic multi-round research across multiple angles, sources, and validation steps before producing a conclusion or report. Do NOT use when a quick ad-hoc search, a single-page summary, or a one-off fact lookup is enough.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | deep-research |
| description | Use when a task requires systematic multi-round research across multiple angles, sources, and validation steps before producing a conclusion or report. Do NOT use when a quick ad-hoc search, a single-page summary, or a one-off fact lookup is enough. |
Turn open-ended research into an auditable loop: plan → broad search → deep read → validation → report.
Use for questions that need multiple angles, multiple sources, and explicit verification. Do not use for quick fact lookup, one-page summary, or casual search.
A completed run must produce:
plan.md with 3-6 tracked subquestionsreports/brief.mdreports/full-report.mdreports/report.htmldeep-research/ via omp html-serve publishstate.json with sources and report pathsDo not declare the report complete if reports/report.html is missing, still contains {{MARKER}} placeholders, or has not been published/attempted with omp html-serve publish.
The HTML is not a generic markdown render — it extracts specific level-2 headings by name. Before writing brief.md / full-report.md you MUST load references/reporting.md and use its exact headings (## 核心结论, ## 关键分歧 / 风险, ## 研究目标, ## 未解决问题, ## 关键来源汇总). Mismatched headings make those blocks render as "Not reported" on a blank-looking page even though the files have content.
After build-report, verify it did NOT return status: incomplete and that empty_sections is empty; then confirm the rendered reports/report.html body contains no Not reported (e.g. grep -c "Not reported" reports/report.html). A status: ok from the CLI alone is not proof the page is populated.
| Phase | Required action | Load |
|---|---|---|
| 0. Clarify | Create plan.md; define subquestions and scope. | references/methodology.md |
| 1. Explore | Map dimensions, terms, entities, and viewpoints. | references/source-strategy.md |
| 2. Deepen | Read high-value sources; write Pre-search Reasoning before each search. | references/methodology.md |
| 3. Validate | Add counterpoints, limits, alternatives, risks, and contradictions. | references/source-strategy.md |
| 4. Synthesize | Emit [Round N Synthesis]; update plan.md; decide continue/stop. | references/stop-criteria.md |
| 5. Report | Write Markdown reports, generate HTML, then publish to html-serve under deep-research/. | references/reporting.md, references/html-reporting.md |
Do not stop early when key full-text sources, opposing views, or limitation evidence are missing.
omp deep-research <subcommand> [args]
| Command | Purpose |
|---|---|
init | Create a workspace: --topic, --slug, --mode. |
build-report | Write reports, render reports/report.html, and update state.json. |
Read references/cli.md before the first CLI call or when parameters are unclear.
After omp deep-research build-report creates reports/report.html, publish the HTML report to the deep-research/ directory:
omp html-serve publish /tmp/report.html --to deep-research/2026-06-13.html --source deep-research --tag research --tag <topic-tag>
In real runs, replace /tmp/report.html with the generated reports/report.html path and use the workspace/report name for the target file, for example:
omp html-serve publish "<workspace>/reports/report.html" --to "deep-research/<workspace-name>.html" --source deep-research --tag research --tag <topic-tag>
Return the localhost_url and tailscale_url. If publishing fails, report the error and still return the local reports/report.html path.
Prefer omp web-operator:
search-multi for 2-3 complementary platforms per roundread-url for full textFallback to WebSearch / WebFetch only when web-operator fails. Do not hand-roll CDP or curl page scraping.
Default workspace root: ~/.local/share/oh-my-superpowers/deep-research/. Override with DEEP_RESEARCH_DATA_DIR. Directory layout and state.json schema live in references/workspace.md.
Generate a daily AI tech brief: scan configured sources (X accounts, blogs, tech communities), filter high-value items, and render a card-based HTML newsletter. Use when the user asks for a daily AI digest, tech briefing, information scan, or says "今日简报", "AI 日报", "tech brief", "扫一遍今天的信息". Do NOT use for deep research on a single topic (use deep-research), or for X account content curation without the brief format.
Generate a daily GitHub trending project-intelligence report: fetch trending repos, study README/API evidence, answer each repo's purpose/content/value in Chinese, and publish a lens-first HTML digest to html-serve. Use when the user asks for GitHub trending, hot repos, 今日 GitHub 热门, trending 项目报告, or a daily open-source project digest. Do NOT use for deep research on a single repo (use deep-research) or general AI news briefs (use daily-ai-brief).
Use when designing a polished static HTML page prototype and its DESIGN.md: clarify the user's page scenario, choose an information organization model, search DESIGN.md references, create multiple adjustable prototypes, publish them to html-serve, and iterate from user feedback. Do NOT use for routine publishing of an already-approved HTML template.
Use when turning one or more IMAP mailboxes into structured local outputs: classified JSONL events, downloaded attachments, and auditable processing state. Use for repeatable mailbox ingestion, classification, and attachment archiving SOPs. Do NOT use for sending mail, MCP email tools, one-off email reading, or full email-client workflows.
Use Web Operator to inspect, navigate, extract, and run site-specific SOP workflows in a local Chrome-family browser. Use when the user needs a workflow on a live Chrome tab, a bundled site workflow script should run, or a page already open in Chrome should be inspected or debugged.
Create a new Agent Skill from a requirement — grounding on what a skill is, clarifying what to build, picking a skill pattern, drafting SKILL.md plus references, and reviewing the result. Use when the user says "写个 skill", "创建 skill", "把这个能力做成 skill", "build a skill", "make a skill", or asks to author or scaffold a skill, even without naming the format. Do NOT use for restructuring an existing skill (skill-refine), auditing a finished skill (skill-review), or non-skill Markdown.