Deep Research

Core Purpose

Deliver citation-backed, verified research reports through a structured pipeline with durable evidence tracking, progressive context management, and resumable markdown-first packaging.

Default report language: Final reports should use Chinese section headings and Chinese narrative by default. Existing English headings remain valid as compatible input for validation and resume.

Autonomy Principle: Operate independently. Infer reasonable assumptions from context. Pause only for critical blockers, contradictory requirements, or missing access.

Non-Silent Degradation Rule: The skill must not silently drop parallel deep-dive retrieval, persona red-team critique, or continuation handoff. If the current runtime cannot satisfy one of these directly, use the local orchestration helper or stop and report the blocker explicitly.

Decision Tree

Request Analysis
+-- Simple lookup? --> STOP: Use a normal browse/search workflow
+-- Debugging or implementation task? --> STOP: Use standard coding/debugging workflow
+-- Complex analysis or synthesis 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 12-24 months unless the prompt sets a wider window.

Workflow Overview

Phase	Name	Quick	Standard	Deep	UltraDeep
1	SCOPE	Y	Y	Y	Y
2	PLAN	-	Y	Y	Y
3	RETRIEVE	Y	Y	Y	Y
4	TRIANGULATE	-	Y	Y	Y
4.5	OUTLINE REFINEMENT	-	Y	Y	Y
5	SYNTHESIZE	-	Y	Y	Y
6	CRITIQUE	-	-	Y	Y
7	REFINE	-	-	Y	Y
8	PACKAGE	Y	Y	Y	Y

Execution

On invocation, load relevant reference files:

Phase 1-7: Load methodology.md for detailed phase instructions
Phase 8 (Report): Load report-assembly.md for progressive generation
Quality checks: Load quality-gates.md
Long reports (>18K words): Load continuation.md
Runtime adapter: Load runtime-codex.md or runtime-opencode.md when running in that environment

Templates:

Report structure: report_template.md

Scripts:

python scripts/research_engine.py --query [topic] --mode [mode] --runtime [codex|opencode|generic]
python scripts/research_engine.py --query [topic] --mode [mode] --runtime [codex|opencode|generic] --skeleton-only
python scripts/research_engine.py --query [topic] --mode [mode] --runtime [codex|opencode|generic] --attempt-autowrite
python scripts/research_engine.py --query [topic] --mode [mode] --runtime [codex|opencode|generic] --auto-continue
python scripts/research_engine.py --resume ./research_[YYYYMMDD]_[topic_slug]/run_state.json --runtime [codex|opencode|generic]
python scripts/research_engine.py --resume ./research_[YYYYMMDD]_[topic_slug]/continuation_state.json --runtime [codex|opencode|generic]
python scripts/validate_report.py --report [path]
python scripts/verify_citations.py --report [path]
python scripts/source_evaluator.py

Non-Negotiable Capabilities

1. Parallel Deep-Dive Retrieval

In standard, deep, and ultradeep modes, retrieval must include 2-3 focused deep-dive tracks beyond broad search.
Preferred execution order: native runtime delegation/subagents -> local orchestration helper -> explicit blocker.
Do not silently collapse to a single shallow retrieval pass.

2. Persona Red-Team Critique

In deep and ultradeep modes, critique must include these personas:
- Skeptical Practitioner
- Adversarial Reviewer
- Implementation Engineer
If critique exposes a critical knowledge gap, run targeted delta queries before final packaging.

3. Continuation Handoff

When the report no longer fits comfortably in one pass, persist continuation_state.json and continue through native runtime support or an explicit helper resume.
Preferred execution order: native runtime continuation/delegation -> local orchestration helper -> explicit blocker.
Do not return a partial report as if it were complete.

Output Contract

research_engine.py should be treated as an orchestration helper, not a fully automatic report writer. In most runtimes it mainly generates phase artifacts, status files, and the report skeleton. Actual retrieval, analysis, prose drafting, and artifact backfill usually still need to be performed by the current agent or delegated subagents.

Required sections:

Executive Summary (50-400 words)
Introduction (scope, methodology, assumptions)
Main Analysis (4-8 findings, 600-2,000 words each, cited)
Synthesis & Insights (patterns, implications)
Limitations & Caveats
Recommendations
Bibliography (COMPLETE - every citation, no placeholders)
Appendix: Methodology

Output directory: ./research_[YYYYMMDD]_[topic_slug]/

Expected files:

report.md - primary deliverable
sources.json - durable source ledger and citation metadata
run_state.json - checkpoint state, section progress, next-section batch, and runtime metadata for resume/orchestration
continuation_state.json - only when the report spans multiple passes
phase_scope.json - capability contract artifact for research framing quality
phase_plan.json - capability contract artifact for research roadmap quality
phase_retrieve.json - capability contract artifact for parallel deep-dive retrieval
phase_triangulate.json - capability contract artifact for claim verification and contradiction tracking
phase_outline_refinement.json - capability contract artifact for evidence-driven outline updates
phase_synthesize.json - capability contract artifact for synthesis quality and argument formation
phase_critique.json - capability contract artifact for persona critique and delta-query loop
phase_refine.json - capability contract artifact for critique follow-up and gap closure

Quality standards:

10+ sources, 3+ per major claim
All factual claims cited immediately as [N]
No placeholders, no fabricated citations
Prose-first (>=80%), bullets used sparingly
deep and ultradeep must include persona-based red-team critique before delivery

Optional helpers and extensions:

source_evaluator.py can be used when the runtime wants explicit source scoring

Environment Assumptions

This skill assumes the current agent environment can do the following:

Search the web or another current-information source
Open and read individual source pages
Run local shell commands and repository scripts
Create and edit files in the working directory
Either provide native delegation/continuation primitives, or allow the skill to call the local orchestration helper

Optional niceties may degrade gracefully. The three non-negotiable capabilities above may not.

When to Use / NOT Use

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

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

deep-research