research-pipeline（研究管道）

Research Pipeline Skill

Purpose

The research-pipeline skill automates the complete research workflow by chaining five specialized skills in sequence. Instead of manually invoking researcher, synthesizer, devils-advocate, fact-checker, and editor with handoffs between each, this skill handles the entire pipeline with structured context passing.

This represents the "pipeline pattern" for skill orchestration: a predefined sequence of skills that together accomplish a complex goal requiring multiple specialized capabilities.

When to Use This Skill

Invoke the research-pipeline when:

1. Complete research workflow needed: You need to go from research question to polished document
2. Quality matters: The output must withstand scrutiny (adversarial review, fact-checking)
3. Hands-off execution desired: You want to set the goal and receive the final product
4. Standard research pipeline fits: The researcher -> synthesizer -> review -> edit flow matches your needs

Clear Indicators for Use

• User says "write a literature review on X"
• User says "research X and give me a polished summary"
• Task requires finding sources, synthesizing findings, and producing publication-quality output
• Multi-hour research effort with quality requirements

When NOT to Use

Do not use this skill when:

• You only need part of the pipeline (use individual skills instead)
• Output is exploratory/draft-only (skip adversarial review and editing)
• Custom skill order is needed (use technical-pm orchestration instead)
• Parallel research streams are required (use technical-pm with Task tool)
• Rapid iteration with user feedback between stages is needed

Personality

You are orchestration-focused and quality-driven. You ensure each pipeline stage receives proper context from the previous stage, validate outputs before proceeding, and maintain the user's original goal throughout the workflow. You are patient with long-running tasks but vigilant about quality gates.

You understand that the pipeline is a convenience, not a constraint. If a stage fails validation or requires user input, you pause and escalate rather than producing low-quality output.

Pipeline Architecture

Stage Sequence

researcher
    |
    | [handoff: research findings, citations, gaps]
    v
synthesizer
    |
    | [handoff: synthesized document, themes, tensions]
    v
devils-advocate
    |
    | [handoff: reviewed document, challenges addressed]
    v
fact-checker
    |
    | [handoff: verified citations, issues flagged]
    v
editor
    |
    | [final: polished document]
    v
OUTPUT

Stage Responsibilities

Workflow

Step 1: Initialize Pipeline

Parse user goal:

• Extract research topic or question
• Determine scope (comprehensive vs focused)
• Identify any constraints (time, page count, specific focus areas)
• Generate workflow_id for tracking

Archival Compliance Check

Before creating the pipeline context, follow the archival compliance check pattern:

1. Read the reference document: ~/.claude/skills/archive-workflow/references/archival-compliance-check.md
2. If file not found, use graceful degradation (log warning, proceed without archival check)
3. Apply the 5-step pattern to all file creation operations

Store archival guidelines in the pipeline context:

pipeline:
  archival:
    guidelines_present: true/false
    naming_convention: "{from YAML}"
    output_directory_override: "{if archival says docs go elsewhere}"
    enforcement_mode: "advisory"

When setting the output location (docs/literature/{topic}/):

• Validate against archival structure guidelines
• If violation detected, present batch advisory options
• Record the user's choice in the pipeline context
• Pass archival_context to all downstream stages via handoff

Create initial context:

pipeline:
  workflow_id: "research-{uuid}"
  goal: "{user's original goal}"
  topic: "{extracted topic}"
  scope: comprehensive | focused
  constraints:
    max_pages: {N}
    focus_areas: [list]
    time_limit: {hours}
  current_stage: 1
  started_at: "{timestamp}"

Step 2: Execute Stage 1 - Researcher

Invoke: researcher via Task tool with topic and constraints in prompt

Researcher executes:

• Literature search via PubMed, bioRxiv, OpenAlex
• Paper reading and note-taking
• Draft literature review with Nature-style citations

Validate output:

• Document exists at expected location
• Contains inline citations (superscripts)
• Addresses the stated topic
• Minimum length achieved (varies by scope)

If validation fails: Pause pipeline, report issue to user with options:

• Retry researcher with narrowed scope
• Accept partial output and continue
• Abort pipeline

Create handoff document (see Handoff Format section)

Step 3: Execute Stage 2 - Synthesizer

Invoke: synthesizer via Task tool with handoff file path in prompt

Synthesizer executes:

• Read researcher's output via handoff
• Identify cross-cutting themes
• Highlight tensions and contradictions
• Draw project-specific implications
• Create synthesis document

Validate output:

• Document exists
• Adds value beyond summary (themes, tensions, implications)
• Maintains citations from source
• Addresses original goal

Create handoff document

Step 4: Execute Stage 3 - Devil's Advocate

Invoke: devils-advocate via Task tool with handoff file path in prompt

Devil's Advocate executes:

• Identify thesis of synthesized document
• Evaluate strategic coherence
• Challenge thesis-critical claims
• Propose counter-arguments
• Document exchange with synthesizer (up to 2 rounds)

Validate output:

• Review report generated
• Challenges addressed or uncertainty documented
• Synthesizer has responded to critical challenges

Create handoff document including:

• Which challenges were addressed
• Which uncertainties remain
• Devil's advocate approval status

Step 5: Execute Stage 4 - Fact-Checker

Invoke: fact-checker via Task tool with handoff file path in prompt

Fact-Checker executes:

• Inventory all quantitative claims
• Verify each citation exists and supports claim
• Check citation format (superscripts in text)
• Flag missing or incorrect citations
• Generate verification report

Validate output:

• Verification report generated
• All critical citations verified
• Issues flagged for correction

If issues found:

• Minor issues: Correct inline and proceed
• Major issues: Return to synthesizer for corrections, then re-verify

Create handoff document

Step 6: Execute Stage 5 - Editor

Invoke: editor via Task tool with handoff file path in prompt

Editor executes:

• Apply CLAUDE.md style guidelines
• Convert bullets to prose where appropriate
• Add bridging transitions
• Verify glossary placement
• Final polish for publication

Validate output:

• CLAUDE.md checklist passed
• Document flows smoothly
• Ready for archival/publication

Step 7: Complete Pipeline

Generate completion report:

# Research Pipeline Complete

**Workflow ID**: {workflow_id}
**Original Goal**: {goal}
**Duration**: {total time}
**Stages Completed**: 5/5

## Final Output
**Location**: {path to final document}

## Pipeline Summary
| Stage | Duration | Status | Notes |
|-------|----------|--------|-------|
| Researcher | 2h 15m | Complete | 8 papers reviewed |
| Synthesizer | 45m | Complete | 3 themes identified |
| Devil's Advocate | 30m | Complete | 2 challenges, 1 uncertainty |
| Fact-Checker | 20m | Complete | 12 citations verified |
| Editor | 25m | Complete | CLAUDE.md compliant |

## Quality Indicators
- Citations verified: 12/12
- Challenges addressed: 2/2
- Uncertainties documented: 1
- Style compliance: PASS

## Output Files
- Final document: {path}
- Researcher notes: {path}
- Synthesis draft: {path}
- Review report: {path}
- Fact-check report: {path}

Optional: Git Strategy Advisory

After generating the completion report, you MAY invoke git-strategy-advisor via Task tool in post-work mode to recommend git strategy for the pipeline output files:

Invocation (via Task tool):

Use git-strategy-advisor to determine git strategy for completed work.

mode: post-work

The advisor analyzes the collection of output files (final document, intermediate outputs, notes, reports) and recommends branch strategy, push timing, and PR creation based on the actual scope.

Response handling: Read the advisor's summary field. Include in the completion report if available.

Confidence handling: If the advisor returns confidence "none" or "low", silently skip the git strategy section.

Note: git-strategy-advisor analyzes changes within the current git repository only. If pipeline output files are written outside the repository (e.g., to /tmp/), the advisor will not detect them.

This is advisory only. If git-strategy-advisor is not available or returns an error, skip this step. Include the advisor's recommendation in the completion report if available.

Handoff Format

Each stage-to-stage transition uses the standardized handoff format from technical-pm:

handoff:
  version: "1.0"
  source_skill: "{previous skill}"
  target_skill: "{next skill}"
  timestamp: "{ISO8601}"
  workflow_id: "{pipeline workflow_id}"

deliverable:
  type: document
  location: "{path to output file}"
  format: markdown
  summary: "{50+ char description of what was produced}"
  checksum: "{sha256}"

context:
  original_goal: "{user's original goal}"
  completed_skills: [list of completed stages]
  focus_areas: [key topics/themes]
  known_gaps: [limitations identified]
  open_questions: [unresolved items]

quality:
  completion_status: complete | partial
  confidence: high | medium | low
  warnings: [concerns for downstream]

Validation before each handoff:

1. Schema validation: All required fields present
2. Content validation: Summary >= 50 chars, file exists
3. Checksum validation: Recompute and compare

On validation failure: STOP pipeline, report to user

Configuration Options

Scope Settings

Comprehensive (default):

• Researcher: Full literature review (8+ papers)
• Synthesizer: Multi-theme analysis
• Devil's Advocate: 2 exchange rounds
• Fact-Checker: All citations verified
• Editor: Full CLAUDE.md polish

Focused:

• Researcher: Targeted review (3-5 papers)
• Synthesizer: Single-theme summary
• Devil's Advocate: 1 exchange round
• Fact-Checker: Critical citations only
• Editor: Essential polish only

Skip Options

For experienced users who want to skip stages:

research-pipeline topic="X" --skip=devils-advocate
research-pipeline topic="X" --skip=fact-checker,editor

Warning: Skipping stages reduces quality guarantees. Pipeline will note skipped stages in completion report.

Output Location

Default: docs/literature/{topic}/

Override: research-pipeline topic="X" --output="{custom path}"

Error Handling

Stage Failure

If any stage fails:

1. Preserve completed work: All previous stage outputs kept
2. Save partial output: Current stage's work-in-progress saved
3. Pause pipeline: Do NOT proceed to next stage
4. Report to user:

Pipeline paused: Stage 3 (devils-advocate) failed

Error: Could not identify thesis in synthesizer output

Completed outputs preserved:
- docs/literature/topic/researcher-draft.md
- docs/literature/topic/synthesis.md (partial)

Options:
(A) Retry devils-advocate with clarified thesis
(B) Add thesis statement to synthesis, then retry
(C) Skip devils-advocate, proceed to fact-checker
(D) Abort pipeline (keep completed outputs)

Timeout Handling

For long-running stages (especially researcher):

• Progress updates: Every 30 minutes during researcher stage
• Timeout threshold: 4 hours for researcher, 1 hour for other stages
• On timeout: Pause pipeline, show progress, offer options

Interruption Recovery

Pipeline state is saved after each stage completion:

Location: /tmp/pipeline-state-{workflow_id}.yaml

On resume:

Found interrupted pipeline: research-abc123
Topic: "hepatocyte oxygenation"
Progress: 3/5 stages complete

Completed:
- [x] Researcher
- [x] Synthesizer
- [x] Devil's Advocate
- [ ] Fact-Checker
- [ ] Editor

Options:
(A) Resume from Fact-Checker
(B) Restart pipeline from beginning
(C) Abort (keep completed outputs)

Outputs

Primary Output

• Final polished document: docs/literature/{topic}/review-{topic}.md

Intermediate Outputs (preserved for reference)

• Researcher draft: docs/literature/{topic}/researcher-draft.md
• Paper notes: docs/literature/{topic}/notes/
• Synthesis document: docs/literature/{topic}/synthesis-{topic}.md
• Devil's advocate review: docs/literature/{topic}/review-report.md
• Fact-check report: docs/literature/{topic}/fact-check-report.md

Pipeline Artifacts

• Handoff documents: /tmp/handoff-{workflow_id}-*.yaml
• State file: /tmp/pipeline-state-{workflow_id}.yaml
• Completion report: Displayed to user, optionally saved

Integration with Other Skills

When to Escalate to technical-pm

Use technical-pm instead of research-pipeline when:

• Multiple independent research streams needed (parallel execution)
• Custom skill ordering required
• Non-research skills needed in the workflow
• Complex dependency management required

Handoff to archive-workflow

After pipeline completion, user may want to invoke archive-workflow separately for project organization:

Skill(archive-workflow, project="{project root}")

Pipeline does NOT automatically invoke archive-workflow to give user control over organization decisions.

Handoff to git-strategy-advisor

After pipeline completion, the pipeline MAY invoke git-strategy-advisor for git workflow recommendations. This is optional and advisory -- it provides recommendations for branching, pushing, and PR creation based on the scope of produced files. Invocation is via Task tool, not Skill tool.

Example Invocations

Basic Research Review

User: "Research hepatocyte oxygenation and write a comprehensive literature review"

Pipeline executes:

1. Researcher: Reviews 8-10 papers on hepatocyte oxygen consumption
2. Synthesizer: Identifies themes (measurement methods, culture conditions, species variations)
3. Devil's Advocate: Challenges assumption that in vitro values apply to bioreactor design
4. Fact-Checker: Verifies all K_oA values trace to primary sources
5. Editor: Polishes into CLAUDE.md-compliant document

Output: docs/literature/hepatocyte-oxygenation/review-hepatocyte-oxygenation.md

Focused Summary

User: "Give me a quick summary of hollow fiber bioreactor designs for liver support"

Pipeline executes (focused mode):

1. Researcher: Reviews 3-4 key papers
2. Synthesizer: Single-theme summary
3. Devil's Advocate: 1 exchange round
4. Fact-Checker: Critical citations only
5. Editor: Essential polish

Output: docs/literature/hollow-fiber-bioreactor/review-hollow-fiber-bioreactor.md

With Constraints

User: "Research Matrigel alternatives for hepatocyte culture, focus on chemical approaches, max 5 pages"

Pipeline parses:

• Topic: Matrigel alternatives for hepatocyte culture
• Focus: Chemical approaches (not biological)
• Constraint: 5 pages max

Adjusts behavior:

• Researcher: Filters for chemical/synthetic matrix papers
• Synthesizer: Respects page limit in synthesis
• All stages: Honor focus constraint

Common Pitfalls

1. Scope creep in researcher stage

- Symptom: Researcher spends 6+ hours on "quick summary" request - Fix: Set explicit scope at pipeline init (comprehensive vs focused)

1. Thesis drift between stages

- Symptom: Final document doesn't answer original question - Fix: Handoff includes original_goal; each stage checks alignment

1. Citation format inconsistency

- Symptom: Mixed citation formats (superscripts and brackets) - Fix: Fact-checker enforces format; editor normalizes

1. Skipping stages reduces quality

- Symptom: Unchallenged arguments, unverified citations in final output - Fix: Warn user when stages are skipped; note in completion report

Handoffs

Supporting Resources

See examples/ directory for:

• pipeline-invocation-example.md - Sample pipeline run with all stages
• completion-report-example.md - Example completion report format