菜单
Produce a sourced research report in the user's language using a structured, evidence-first workflow. Supports quick, standard, deep, and exhaustive research modes. Prioritizes evidence quality, entity disambiguation, claim verification, contradiction checking, uncertainty handling, and clear synthesis before final writing.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Plan, review, and then assemble a stakeholder-ready dashboard package from the current analysis run, with a native interactive canvas experience.
Data analysis skill hub. Routes to the right specialist subskill depending on the request — exploration, query writing, end-to-end analysis, visualization, validation, dashboard planning plus assembly, or recurring snapshot refresh.
End-to-end data analysis — from a quick metric lookup to a formal stakeholder report. Combines schema discovery, SQL execution, optional visualization, and an evidence-based summary. Use for general "analyze / what's happening with…" requests.
Create, edit, redline, comment on, and visually verify `.docx`, Word, and Google Docs-targeted document artifacts. Use for Word documents, `.docx` files, document creation, document edits, tracked changes, comments, render QA, and Google Docs-ready local DOCX generation.
Create stunning, animation-rich browser-native HTML presentations. Use only when the user explicitly wants an HTML/web presentation or animated interactive browser deck. Do not use for `.ppt`, `.pptx`, PowerPoint, Google Slides, native deck creation, native deck editing, deck templates, or PPTX conversion/export requests; use the `pptx` skill for those.
Work with PDF documents safely, including querying existing PDFs and creating simple image-based PDFs from workspace images.
| name | research |
| description | Produce a sourced research report in the user's language using a structured, evidence-first workflow. Supports quick, standard, deep, and exhaustive research modes. Prioritizes evidence quality, entity disambiguation, claim verification, contradiction checking, uncertainty handling, and clear synthesis before final writing. |
| tools | ["google_search","request_plan_approval","request_clarification","request_human_action"] |
| requires_hitl_plan | true |
| requires_workspace_artifacts | true |
| required_artifacts_mode | full_pack |
| required_artifacts | ["/question.txt","/research_mode.md","/entity_disambiguation.md","/preliminary_search_notes.md","/research_plan.md","/source_register.md","/research_notes.md","/claim_evidence_matrix.md","/contradictions_and_uncertainties.md","/red_flags_and_exclusions.md","/knowledge_graph.md","/synthesis.md","/final_report_audit.md","/final_quality_check.md","/final-research-report.md"] |
| pre_plan_search_limit | 3 |
| source_skills | ["research-core","research-sub_researcher","research-critique"] |
Use this skill to research a topic and deliver a structured, well-sourced report in the same language as the user.
This skill emphasizes:
The goal is not merely to produce a long report. The goal is to produce a report whose major claims have earned the right to appear.
Before full research begins, classify the request into one of these modes and record it in /research_mode.md.
Use when the user asks for a narrow factual answer or a short explanation.
/final-research-report.mdUse for normal explanatory or comparative research.
/final-research-report.mdUse when the query involves:
Requirements:
/final-research-report.mdUse when the user explicitly asks for an exhaustive landscape, investment memo, policy review, academic-style report, or due-diligence report.
/final-research-report.mdIf the user does not specify a depth, choose the appropriate mode automatically and explain the choice briefly in the research plan.
Unless explicitly stated otherwise, "final report word count" refers only to the words in /final-research-report.md.
Do not count:
/claim_evidence_matrix.md/research_notes.md/comparison_matrix.md/knowledge_graph.md/synthesis.md/source_register.md/final_quality_check.mdSupporting artifacts may be longer than the final report, but they do not count toward the final report word-count requirement.
Never report an estimated final report word count. Always compute the count from the actual final report file.
The final report's Word Count section must use this exact form:
## Word Count
1234 words
Use the actual integer count. Do not add qualifiers, parentheses, methodology notes, estimates, or explanations in the final report.
Forbidden final-report word-count phrases:
If any of these phrases appear in the final report's Word Count section, the report fails the completion gate.
Known failing examples:
## Word Count
(Manual count of /final-research-report.md content)
Approx 6,400 words.
## Word Count
Approximately 7,200 words.
Both examples fail because they are not computed integer counts and contain forbidden qualifiers.
Write the original user question to /question.txt.
Do not reinterpret the question silently.
Create /research_mode.md.
Include:
Example:
# Research Mode
Selected mode: Deep Research
Reason:
The user is comparing multiple entities and asks for technical and strategic evaluation.
Expected final report length:
6,000-10,000 words
Minimum sources:
18 credible sources
Minimum primary sources:
8 where available
Create /entity_disambiguation.md.
For every named entity, product, company, model, framework, law, or technical term in the user question:
Rules:
request_clarification.Example:
# Entity Disambiguation
## Entity: Pi Agent
Possible meanings:
1. Inflection AI's Pi assistant
2. A coding-agent project using the name Pi
3. A component inside another framework
Decision:
The report will distinguish these meanings unless primary sources prove they are the same.
Clarification needed:
No, because the ambiguity itself is relevant to the research.
Run 2-3 lightweight searches before plan approval.
Purpose:
Record findings in /preliminary_search_notes.md.
Include:
Do not perform full evidence gathering yet.
Create /research_plan.md.
The plan must include:
Then call request_plan_approval with:
plan_titleplan_summaryexecution_checklistrisky_actions when relevantreviewer_feedback when editing a previously proposed planA limited pre-plan search of 2-3 calls is allowed before approval, but the main evidence gathering must wait for approval.
If the user edits the plan, update /research_plan.md and request approval again.
If the user rejects the plan, stop.
Use google_search to collect sources.
Prioritize primary sources first.
Tier A - Primary Sources:
Tier B - Reputable Secondary Sources:
Tier C - Low-Confidence Sources:
Rules:
Create /source_register.md.
For every source, include:
## Source [1]
Title:
URL:
Final URL after redirects:
Publisher / author:
Date:
Access date:
Fetch status: reachable / unreachable / paywalled / redirected / failed
Source tier: A / B / C
Primary or secondary:
Key claims:
Claims supported:
Claims not supported:
Useful for:
Limitations:
Reliability notes:
The final report may cite only sources that appear in the source register.
Before finalizing, open every URL listed in the final report's Sources section.
For each cited source, record in /source_register.md:
Rules:
Create /research_notes.md.
Include:
No report prose should begin before this file exists.
Create /claim_evidence_matrix.md.
Every major factual claim must be recorded before appearing in the final report.
Use this structure:
| Claim | Source(s) | Source tier | Evidence summary | Confidence | Include? |
|---|---|---|---|---|---|
| Claim text | [1], [2] | A / B | Short evidence note | High / Medium / Low / Exclude | Yes / No |
Confidence rules:
Rules:
Create /contradictions_and_uncertainties.md.
For each major claim, perform contradiction-oriented searches where appropriate.
Useful search patterns:
[claim keywords] official
[claim keywords] GitHub
[claim keywords] documentation
[claim keywords] benchmark
[claim keywords] pricing
[claim keywords] CVE
[claim keywords] controversy
[claim keywords] fake
[claim keywords] not true
[claim keywords] acquisition
[claim keywords] release notes
Record:
Create /red_flags_and_exclusions.md.
Use this file to track claims that were tempting but should not be included.
Include:
| Excluded claim | Reason excluded | Source issue | Safer wording |
|---|---|---|---|
Claims that require caution include:
For Standard, Deep, and Exhaustive comparative reports, create one standalone dossier per major entity.
Use filenames such as:
/01_entity_openclaw.md
/02_entity_hermes.md
/03_entity_pi.md
Each dossier should include:
For Deep Research mode, each dossier should be 800-1,500 words.
For Exhaustive Research mode, dossiers may be longer.
Write each major section as a separate Markdown file using zero-padded filenames:
/01_tldr.md
/02_scope.md
/03_methodology.md
/04_background.md
/05_findings.md
/06_comparative_analysis.md
/07_implications.md
/08_recommendations.md
/09_conclusion.md
Rules:
Create /knowledge_graph.md.
Use Mermaid.
The graph must include:
Example:
graph TD
A[Company A] -->|develops| B[Product B]
B -->|competes with| C[Product C]
B -->|depends on| D[Model Layer]
C -->|targets| E[Enterprise Users]
The knowledge graph should reflect the actual evidence, not generic grouping.
For comparative research, create /comparison_matrix.md before final synthesis.
Use dimensions relevant to the topic.
For technical or product comparisons, include:
Create /synthesis.md.
This file must exist before the final report.
Include:
The synthesis must go beyond summarizing sources. Explain what the evidence means.
Create /final-research-report.md, unless the user specifies another filename.
The final report must be based on the section files, synthesis file, comparison matrix when present, and claim-evidence matrix.
The final report must itself satisfy the selected mode's word-count range. Supporting artifacts are evidence and drafting material; they are not part of the final report word count.
Do not merely stitch sections mechanically.
The final consolidation must:
/final-research-report.md after writing; do not estimate itAdapt as needed, but default to:
# Title
## TL;DR
## Scope and Research Questions
## Methodology and Evidence Quality
## Entity Disambiguation
## Background
## Findings
## Comparative Analysis
## Risks and Limitations
## Strategic or Practical Implications
## Recommendations
## Limits of Evidence
## Conclusion
## Sources
## Word Count
For technical comparisons, include:
## Model Layer vs Agent Layer
## Architecture and Runtime
## Memory and State
## Tool Use and Integrations
## Deployment and Operations
## Security Posture
## Use-Case Fit
For policy topics, include:
## Regulatory Context
## Stakeholders
## Compliance Implications
## Enforcement Risks
## Alternative Interpretations
For market topics, include:
## Market Structure
## Competitive Landscape
## Business Model
## Adoption Signals
## Risks to the Thesis
Use numeric citations:
This is a factual claim [1].
Rules:
End with:
## Sources
[1] Source Title: URL
[2] Source Title: URL
Use a clear, professional tone.
Avoid hype.
Do not use these unless directly supported by strong evidence:
Prefer:
If evidence is limited, include a Limits of Evidence section.
If sources conflict, include a Conflicting Evidence subsection.
If primary sources cannot be found, say so clearly.
If the topic is speculative or emerging, state that clearly.
Never fabricate missing data to meet word count.
Never fill gaps with confident-sounding assumptions.
If /final-research-report.md is below the selected mode minimum:
For Deep Research mode, if /final-research-report.md is below 6,000 words, it fails the quality gate even if all supporting artifacts combined exceed 6,000 words.
For Exhaustive Research mode, if /final-research-report.md is below 10,000 words, it fails the quality gate even if all supporting artifacts combined exceed 10,000 words.
Before finalizing, compute the actual word count of /final-research-report.md programmatically. Prefer a deterministic file-based count such as wc -w final-research-report.md from the workspace root.
Write this to /final_quality_check.md:
## Mandatory Computed Word Count
- Final report file: `/final-research-report.md`
- Computed word count:
- Required word count range:
- Pass / fail:
The computed word count must be derived from the final report file only. Do not estimate it. Do not use manual estimates. Do not count workspace artifacts, notes, dossiers, matrices, source registers, quality checks, or chat messages.
If the computed word count is below the selected mode minimum, do not finalize. Expand the final report using the existing dossiers, synthesis, comparison matrix, source register, research notes, and claim-evidence matrix, then recompute the word count.
Before declaring completion, create /final_report_audit.md.
This audit must inspect /final-research-report.md itself, not the artifact pack.
Include:
# Final Report Audit
Final report path:
Computed final report word count:
Required range:
Word count command or method:
Word Count section exact text:
Forbidden word-count phrase present: yes / no
Sources listed in final report:
Source URLs opened:
Unreachable sources:
Sources missing from source register:
High-risk uncited claims found:
Low-confidence claims outside uncertainty sections:
Excluded claims found in final report:
Completion decision: pass / fail
The audit fails if the final report contains:
(Manual count of /final-research-report.md content)Approxapproximatelyestimated/source_register.mdIf /final_report_audit.md says Completion decision: fail, do not finalize and do not claim completion in chat.
Before declaring completion, produce /final_quality_check.md with:
/final-research-report.md/claim_evidence_matrix.md/final_report_audit.md completion decisionCompletion is forbidden if:
/source_register.md.words./final_report_audit.md is missing or says Completion decision: fail.Before finalizing, run a self-check and record it in /final_quality_check.md.
Check:
# Final Quality Check
- [ ] Research mode matches query complexity
- [ ] Entity disambiguation completed
- [ ] Source register completed
- [ ] Claim-evidence matrix completed
- [ ] Contradiction search completed
- [ ] Red flags and exclusions recorded
- [ ] Every cited source URL was opened and recorded with fetch status in `/source_register.md`
- [ ] Unreachable URLs are removed or clearly marked unavailable
- [ ] Knowledge graph completed
- [ ] Synthesis completed
- [ ] Mandatory computed word count completed from `/final-research-report.md` only
- [ ] Computed final report word count is within the selected mode range
- [ ] Final report Word Count section does not use approximate/manual/estimated wording
- [ ] Final report Word Count section is exactly an integer followed by `words`
- [ ] `/final_report_audit.md` exists and says `Completion decision: pass`
- [ ] Final report completion gate passed
- [ ] Final report cites only registered sources
- [ ] Unsupported claims removed
- [ ] Low-confidence claims appear only in uncertainty or limits-of-evidence sections
- [ ] Excluded claims do not appear in the final report
- [ ] Uncertainty clearly marked
- [ ] No fake citations
- [ ] No unsupported numbers
- [ ] No unsupported superlatives
- [ ] Report length claim, if any, matches the computed word count
Do not finalize the report until this check is complete.