Deep Research

深度研究 · Deep Research

Core Purpose

Deliver citation-tracked research reports through a structured pipeline with evidence persistence, source identity management, claim-level verification, and progressive context management.

Autonomy Principle: Operate independently. Infer assumptions from context. Only stop for critical errors or incomprehensible queries. Surface high-materiality assumptions explicitly in the Introduction and Methodology rather than silently defaulting.

Dependencies

• lov-dev-blog owns the Skill Publisher website blog publishing contract. deep-research owns research generation and verification; final publishing to blog_posts must use the dev-blog automation semantics.

Decision Tree

Request Analysis
+-- Simple lookup? --> STOP: Use WebSearch
+-- Debugging? --> STOP: Use standard tools
+-- Complex analysis needed? --> CONTINUE

Mode Selection
+-- Initial exploration --> quick (3 phases, 2-5 min)
+-- Standard research --> standard (6 phases, 5-10 min) [DEFAULT]
+-- Critical decision --> deep (8 phases, 10-20 min)
+-- Comprehensive review --> ultradeep (8+ phases, 20-45 min)

Default assumptions: Technical query = technical audience. Comparison = balanced perspective. Trend = recent 1-2 years.

Workflow Overview

Note: Phases 3-5 operate as an evidence loop per section (retrieve → evidence store → refine outline → draft → verify claims → delta-retrieve if needed), not as strict sequential gates.

Execution

On invocation, load relevant reference files:

1. Phase 1-7: Load methodology.md for detailed phase instructions
2. Phase 8 (Report): Load report-assembly.md for progressive generation
3. HTML/PDF output: Load html-generation.md
4. Quality checks: Load quality-gates.md
5. Implementation/tooling research: Load open-source-solutions.md
6. Comparison/selection/adoption research: Load decision-guides.md
7. Long reports (>18K words): Load continuation.md

Templates:

• Report structure: report_template.md
• HTML styling: mckinsey_report_template.html

Scripts:

• python scripts/validate_report.py --report [path]
• python scripts/verify_citations.py --report [path]
• python scripts/md_to_html.py [markdown_path]
• python scripts/validate_open_source_solutions.py --artifact [open_source_solutions.jsonl] --report [report.md] --strict
• python scripts/validate_decision_guide.py --report [report.md] --strict

Post-report publishing (Skill Publisher):

• This is a mandatory completion gate, not an optional follow-up. Do not send the final answer for a publishable report until either the sync command has succeeded or a concrete sync error has been reported.
• After Phase 8 successfully generates the Markdown report, automatically publish it to the Skill Publisher blog system unless the user explicitly says "private", "do not publish", "不要发布", or equivalent.
• This gate depends on lov-dev-blog. The website sync command below is the dev-blog publishing contract for research-origin artifacts.
• Use the generated Markdown file path as the source of truth: cd ${SKILL_WORKSPACE_ROOT}/coding/web && pnpm run sync:research -- [markdown_path]
• If multiple Markdown reports were generated or the exact Markdown path is uncertain, run: cd ${SKILL_WORKSPACE_ROOT}/coding/web && pnpm run sync:research -- --limit 5
• Publishing semantics are owned by lov-dev-blog and executed by the website sync script:
  • New reports are public detail pages (is_visible=true).
  • New reports appear in the /blog index by default (show_in_index=true).
  • Re-syncing an existing report also promotes it into the index unless explicitly hidden.
  • Published reports should carry a cover; the website sync script may auto-generate and upload one when the Markdown artifact does not provide a cover URL.
• Tell the user the final public URL in the form: https://lovstudio.ai/blog/[slug]
• In the final answer, include a one-line publishing status: Published to Skill Publisher: yes/no, plus the public URL when yes.
• If the sync command fails because the website path, environment, or database schema is unavailable, keep the completed research artifacts and surface the exact sync error plus the command to rerun.

Output Contract

Required sections:

• Executive Summary (200-400 words)
• Introduction (scope, methodology, assumptions)
• Main Analysis (4-8 findings, 600-2,000 words each, cited)
• Decision Guide (required for comparison, selection, procurement, architecture-choice, or adoption reports; place before detailed findings)
• Open-Source Solutions Landscape (required when the topic concerns software, tooling, automation, implementation, or deployable solutions)
• Synthesis & Insights (patterns, implications)
• Limitations & Caveats
• Recommendations
• Bibliography (COMPLETE - every citation, no placeholders)
• Methodology Appendix

Output files (all to ~/Documents/[Topic]_Research_[YYYYMMDD]/):

• Markdown (primary source of truth)
• sources.jsonl — stable source registry with canonical IDs
• evidence.jsonl — append-only evidence store with quotes and locators
• claims.jsonl — atomic claim ledger with support status
• run_manifest.json — query, mode, assumptions, provider config
• open_source_solutions.jsonl — canonical repository registry for applicable implementation/tooling research; one verified repository per line
• HTML (McKinsey style, auto-opened)
• PDF (professional print, auto-opened)

Quality standards:

• 10+ sources, 3+ per major claim (cluster-independent, not just count)
• All factual claims cited immediately [N] with evidence backing in evidence.jsonl
• Claim-support verification mandatory: no unsupported factual claims pass delivery
• Applicable comparison/selection reports must turn decisive constraints into a branching decision flow with explicit terminal recommendations and a textual fallback; score tables alone do not pass
• Applicable implementation/tooling reports must search GitHub plus other relevant forges, inspect repository evidence beyond README claims, publish a linked comparison table, and persist open_source_solutions.jsonl; an explicit no-results record is required when no repository qualifies
• No placeholders, no fabricated citations
• Prose-first (>=80%), bullets sparingly

When to Use / NOT Use

Use: Comprehensive analysis, technology comparisons, state-of-the-art reviews, multi-perspective investigation, market analysis.

Do NOT use: Simple lookups, debugging, 1-2 search answers, quick time-sensitive queries.

Runtime context (shared)

运行前读取本 Skill 包的 skill.yaml，由宿主提供 skill-runtime/v1 上下文。字段解析顺序为：当前请求、项目上下文、个人 Preferences、品牌 Profile、通用默认值。

• 只使用 Manifest 声明的字段；Profile 保存公开品牌事实，Preferences 保存个人工作偏好。
• required: true 字段缺失时，按 Manifest 的问题配置向用户提出一个聚焦问题；用户明确同意后再保存回答。
• 报错提供可复制的 context_id、字段路径与来源，诊断内容避开秘密、完整私人路径和原始配置。

通用反馈闭环

用户在 Skill 驱动任务中提出修改意见时，继续当前产物前必须执行：

1. 先判断意见是 task-specific（仅本次）还是 reusable（可跨任务复用）。
2. task-specific 只修改当前任务，不改 Skill。
3. reusable 先确定作用域：领域规则先更新对应 canonical Skill；适用于所有 Skill 的规则先更新共享规范。
4. 完成规则更新、版本、lint 与分发核验后，再把修改应用到当前任务。
5. reusable 修改会使此前的“确认”“继续”“发吧”失效；完成当前产物修改和回读后必须停下，等待用户下一步指示，不自动进入发布、提交或其他外部写入。