← Về thư mục

---

name: research-pipeline description: "Multi-phase academic and policy research pipeline with manual gating. Use whenever a research task requires systematic literature gathering, multi-agent deep research, peer review critique, or a structured deliverable (report, gap analysis, policy brief). Triggers on: 'research pipeline', 'deep research', 'literature review', 'policy analysis', 'comparative study', 'R&D task', 'Phase 1/2/3/4' research requests." version: 4.0.0 user-invocable: true allowed-tools: terminal, read_file, write_file, web_search, clarify, delegate_task

---

Research Pipeline — ARS 4.0

Standardized workflow for multi-phase research with strict manual gates and role separation. Domain-agnostic — applies to education policy, tourism, market analysis, institutional benchmarking, or any systematic knowledge production task.

Core Principles

• Manual Gate: No phase auto-advances. CEO must approve each output before the next phase starts.
• Role Separation: Three roles, no overlap — r-and-d (gather + draft), writers (synthesize + edit), reviewer (critique + QA).
• Source Integrity: All primary sources downloaded to sources/, cited APA 7, permalinked via Tailscale Funnel.
• Attribution Discipline: Policy claims must identify the responsible actor (ministry, government, institution) — never conflate levels.
• No Hardcoding: Skills must not reference specific countries, brands, or organizations in logic — only in task context passed at runtime.

---

Phase Map

Phase 1a  r-and-d    Context map + research scope definition
Phase 1b  r-and-d    Source discovery + template design
Phase 2   r-and-d    Deep research (parallel agents per item)
Phase 3   writers    Synthesis + comparative analysis
Phase 4a  reviewer   Peer critique (5-role panel)
Phase 4b  writers    Revision per critique → final deliverable

Each phase ends with: kanban_block(reason="Awaiting CEO gate") + assignee → CEO.

---

Phase 1a — Context Map & Scope

Goal: Establish what is known, what is unknown, what needs to be researched.

1. Read any existing KB or prior research files in the project folder.
2. Identify: research question, sub-questions, target deliverable.
3. List known sources already in hand vs. sources to find.
4. Output: context_map.md — includes gap analysis, 3–5 research hypotheses.
5. Block + report to CEO.

---

Phase 1b — Source Discovery & Template

Goal: Find, download, and organize primary sources. Design output template.

1. Search: academic DBs (arXiv, Semantic Scholar, CrossRef), government portals, official reports.
2. Download PDFs → sources/<domain>/. Run chmod 644 on files, chmod 755 on dirs.
3. Upload key sources to NotebookLM notebook (if available).
4. Design research_output_template.md — all Phase 2 agents use this template.
5. Output: populated sources/ folder + template file.
6. Block + report to CEO.

Source quality hierarchy: Primary (official text) > Peer-reviewed > Institutional report > Secondary analysis. Flag any source not verified.

---

Phase 2 — Deep Research (Parallel Agents)

Goal: One agent per research item/country/topic. All run in parallel via delegate_task.

Each subagent: 1. Loads context from context_map.md + research_output_template.md. 2. Queries NotebookLM for source-grounded answers. 3. Supplements with web search for recent data. 4. Writes output to master_reports/<item>.md following the template exactly. 5. APA 7 citations inline. Flag uncertain data as [unverified].

Parent agent waits for all subagents, then aggregates into index.md. Block + report to CEO.

---

Phase 3 — Synthesis & Comparative Analysis

Goal: writers reads all Phase 2 outputs → single coherent deliverable.

1. Read all master_reports/*.md files.
2. Identify cross-cutting themes, contradictions, data gaps.
3. Write comparative_analysis.md or policy_brief.md depending on deliverable type.
4. Update index.md with links to all outputs.
5. Block + report to CEO.

---

Phase 4a — Peer Critique (5-Role Panel)

Goal: reviewer applies structured critique before CEO sees final output.

Load skill research-critism and apply all 5 reviewer roles:

Role	Focus
Editor in Chief	Research question sharpness, confirmation bias, scope
Methodology Reviewer	Framework appropriateness, logical gaps, DV clarity
Domain Reviewer	Factual accuracy, classification errors, missed evidence
Perspective Reviewer	Institutional incentives, political economy, feasibility
Devil's Advocate	Strongest counter-argument, attribution errors, what the analysis got wrong

Output: grill_report_<N>.md with Editorial Decision: Accept / Minor Revision / Major Revision / Reject + blocking items list.

---

Phase 4b — Revision

Goal: writers addresses all blocking items from 4a critique.

1. Read grill_report_<N>.md — address each blocking item explicitly.
2. Update the deliverable in-place (do NOT create a new file).
3. Write revision log at bottom of grill_report_<N>.md.
4. Block + report to CEO for final approval.

---

Orchestration Rules

• Kanban task per phase: Create before starting, block when done.
• No auto-complete: Only CEO calls kanban_complete.
• Cross-department handoff: Use kanban_block + change assignee only — do not create new tasks.
• Parallel execution cap: Max 3 concurrent subagents per delegate_task batch.
• File permissions: After every write, run chmod 644 <file> && chmod 755 <dir>.
• Funnel links: All deliverable links use https://<funnel-host>/files/<absolute-path>.

---

Folder Structure

content/project/<topic>/
├── context_map.md
├── kb_strategy.md
├── research_output_template.md
├── index.md
├── sources/
│   ├── legal/
│   ├── academic/
│   └── institutional/
├── master_reports/
│   └── <item>.md
├── comparative_analysis.md  (or policy_brief.md)
├── notes/
│   └── grill_report_<N>.md
└── master_reports/

---

Pitfalls

• Confirmation bias: Always ask "what if the core assumption is wrong?" before Phase 2.
• Attribution error: State which actor is responsible for each policy gap — never conflate ministry, government, and legislature.
• Lagging vs. leading indicators: Distinguish explicitly — a metric measuring past outcomes cannot substitute for a forward-looking signal.
• Framework mismatch: When borrowing international frameworks (ILO, OECD, WB), document: (a) original design context, (b) adaptation rationale for local conditions.
• Self-reporting risk: Any data that stakeholders produce about themselves requires independent verification or explicit [self-reported, unverified] flag.