📄 / / root / .hermes / skills / research / policy-analyst / SKILL.md

Policy Analyst & Advocate

name: policy-analyst description: "Policy analysis, evaluation, design, and advocacy. Use when the task involves reading legal/regulatory texts, comparing law vs. implementing regulation, evaluating policy effectiveness against international frameworks (ILO, OECD, World Bank), identifying institutional gaps, or drafting evidence-based policy recommendations. Triggers on: 'policy analysis', 'policy evaluation', 'regulatory gap', 'law vs regulation', 'policy brief', 'institutional reform', 'compare with ILO/OECD/WB', 'vận động chính sách', 'phân tích chính sách', 'RIA'." version: 3.0.0 user-invocable: true allowed-tools: terminal, read_file, web_search, web_extract, clarify

Policy Analyst & Advocate

Domain-agnostic policy analysis framework. Applies to education, labor, health, environment, trade, or any regulatory domain.

For country-specific institutional context (especially Vietnam), load the corresponding domain file from references/domains/ immediately.

1. Domain Directory

Load the relevant domain file immediately when the task matches. Do not wait until after analysis begins.

Domain	Resource Path	When to load
Vietnam Institutional Context	`references/domains/vn.md`	Analyzing Vietnamese laws, circulars, decrees, VĐCS in Vietnam
Education Policy (Context)	`references/domains/education.md`	Education system reform, enrollment quotas, curriculum, teacher policy, school governance
Education Policy (Evaluation)	`references/domains/education-evaluation.md`	Evaluating education policies — Theory of Change, RIA, Ex-ante/Ex-post, OECD framework
ASEAN Education Benchmarks	`references/domains/asean-education-benchmarks.md`	Benchmarking higher education enrollment policies, supply/demand ratios, forward-looking mechanisms, and country profiles in ASEAN/Asia-Pacific
International Frameworks	Inside this SKILL.md	Default baseline for comparative policy, ILO/OECD/WB benchmarking

Domain update rule: When adding a new policy sector (health, labor, environment...), create a new file under references/domains/<sector>.md and add a row to this table. Do not edit the core SKILL.md logic.

2. Four Modes

Mode	When to use
Analyze	Read and decode legal/regulatory text — structure, definitions, operative clauses, omissions
Evaluate	Benchmark policy against an international framework (ILO, OECD, WB, UN SDGs)
Design	Propose concrete, implementable mechanism changes with low institutional cost
Advocate	Build a stakeholder-facing argument — evidence chain, counter-arguments addressed, feasible ask

3. Mode 1 — Analyze: Legal Text Decoding

Goal: Extract operative logic from regulatory documents.

Identify: scope, definitions, operative clauses, enforcement mechanisms, sunset clauses.
Map the 2-tier hierarchy: parent law → implementing regulation. For each key principle in the parent law, check whether the implementing regulation has a corresponding operative mechanism.
Flag: (a) principles stated in the law but not operationalized; (b) mechanisms in the regulation that contradict or narrow the law; (c) undefined terms that carry significant policy weight.
Output table: Law Article | Law Principle | Regulation Clause | Status (Implemented / Narrowed / Absent / Contradicted).

Key question to always ask: Is this stated as a binding obligation or an aspirational principle? Who enforces it and how?

4. Mode 2 — Evaluate: Framework Benchmarking

Goal: Assess whether a policy achieves its stated objectives using an established framework.

Framework selection

ILO: Labor market, skills forecasting, employment quality, LMIS design.
OECD: Education systems, governance, fiscal policy, regulatory quality.
World Bank: Investment climate, institutional capacity, social protection.
UN SDGs: Cross-sectoral sustainability and equity goals.

Steps

Identify which framework is appropriate — document why (design context, country applicability, data requirements).
For frameworks designed for high-formality contexts: explicitly assess applicability to the local informal economy share. Adjust or discard criteria that require data that does not exist.
Map each framework criterion → policy provision → evidence of implementation → gap.
Classify gaps: Structural (missing mechanism), Data (metric exists but no collection system), Institutional (mechanism exists but disconnected from decision-making), Temporal (lagging indicator used where leading indicator needed).

5. Mode 3 — Design: Policy Mechanism Proposals

Goal: Propose specific, implementable changes — not ideal solutions, but feasible ones.

5.1 Institutional Red Lines (Lằn ranh đỏ thể chế)

Before finalizing any policy recommendation, you must identify and map the Institutional Red Lines—existing statutory, budgetary, or jurisdictional boundaries that the recommendation must not violate unless explicitly designed as a high-level structural reform. Crossing these lines without a corresponding legal/budgetary bypass mechanism makes a proposal non-feasible.

For Vietnam Higher Education / Policy context, key red lines include: 1. Institutional Autonomy (Quyền tự chủ): Law 34/2018/QH14 grants universities the statutory right to self-determine enrollment quotas. Recommendations cannot use administrative commands to force quotas or pilot programs on universities. 2. Financial Autonomy (Tự chủ tài chính): Decree 60/2021/NĐ-CP requires universities to balance their budgets. Recommendations for expensive high-tech training (e.g., semiconductors, AI) cannot assume universities will invest in CAPEX (labs, specialized teachers) without matching state funding or PPP frameworks. 3. Jurisdictional Boundaries (Thẩm quyền liên bộ): A single ministry (e.g., MOET) cannot mandate actions or data sharing on another ministry (e.g., MOLISA, Social Insurance) without an overriding Government Decree or Prime Minister Decision.

5.2 Design Steps

For each proposed change: 1. Problem statement: What specific gap does this address? (Reference Mode 1 or 2 output.) 2. Mechanism: Exact change to regulation — which article, what text, what new requirement. 3. Data source: What existing data can power this mechanism? Prefer data already collected by existing agencies (tax, social insurance, employment records) over requiring new data collection. 4. Institutional owner: Which agency is responsible? Is this within their current mandate or requires new mandate? 5. Cost estimate: Administrative burden, compliance cost, transition time. 6. Risk: What could go wrong? Who might resist and why?

Feasibility ladder (choose the lowest rung that solves the problem): - L1: Reuse existing administrative data (social insurance DB, tax records). - L2: Add a reporting field to an existing system. - L3: Mandate data sharing between existing agencies. - L4: Build a new data collection instrument. - L5: Build a new institution.

Start at L1. Only escalate when lower rungs are provably insufficient.

6. Mode 4 — Advocate: Building the Policy Argument

Goal: Construct a persuasive, evidence-based case for a specific policy change.

Argument structure

Framing: State the problem in terms the target audience cares about (economic cost, equity impact, legal inconsistency — choose the frame that resonates).
Evidence chain: Each claim linked to a verifiable source. Flag [estimated] or [self-reported] where data quality is weak.
Attribution discipline: Assign each gap to the correct responsible actor. Do not blame implementer for a legislative gap, or blame legislature for an agency's administrative choice.
Pre-empt counter-arguments: Address the 3 strongest objections before the audience raises them.
The ask: One specific, actionable recommendation per audience. Avoid omnibus proposals.

Attribution levels — always distinguish

Level	Examples
Legislative	Parliament, National Assembly — enacted law
Executive/Regulatory	Ministry — implementing regulations, circulars
Coordination	Cross-ministry mechanism, PM office
Agency	Bureau, department — operational procedures

7. Quality Gates (apply before any deliverable)

[ ] Every claim has a source. Uncited claims marked [assertion — needs evidence].
[ ] International framework applicability is documented, not assumed.
[ ] Attribution is correct — each gap is assigned to the actor who can fix it.
[ ] At least one counter-argument addressed per major recommendation.
[ ] All proposed mechanisms use the lowest feasible rung on the Feasibility Ladder.
[ ] Glossary included for all technical terms (especially English acronyms).
[ ] Formatting & Legibility: Output must be written with short paragraphs, clear section breaks (using ---), and bulleted or numbered sub-points. Avoid giant wall-of-text blocks. All bullet points must have distinct indentation levels to ensure readability on mobile and dark-themed screens.

8. Pitfalls

Formatting density (NEW): Producing dense, unformatted blocks of text for policy design makes the output unreadable on mobile/dark-mode platforms. Always structure text into short paragraph blocks (2-3 sentences max) and use sub-bullets with explicit indents instead of packing multi-level lists into a single paragraph.
Framework worship: Citing ILO/OECD/WB as authority without verifying the framework was designed for the applicable context.
Attribution error: Criticizing a ministry for a gap that is actually a legislative design choice, or vice versa. Refined rule (from TT34 case): When a gap is structural (cross-ministry), attribution must answer TWO questions: (a) which actor bears final accountability, AND (b) what could the implementing ministry have done within its current mandate but did not (e.g. proposing a joint ministerial decision, invoking a higher-level resolution to request data sharing). Never declare a ministry fully exculpated without checking the second question.
Aspirational conflation: Treating a stated policy goal as evidence of an operating mechanism.
Feasibility gap: Recommending data-driven mechanisms in contexts where the required data does not exist or is not reliable.
Confirmation bias: Selecting evidence that supports the initial hypothesis. Always run Devil's Advocate before finalizing recommendations.
LMIS applicability in high-informality contexts: ILO LMIS framework assumes a formal labor market with registration, contracts, and administrative data. In economies with >50% informal employment (e.g. Vietnam >62%), direct application overstates available data. Always check: does the country have BHXH/social insurance coverage, linked tax+employment records, or household survey infrastructure? If not, descend to proxy indicators (web-scraped job ads, household surveys, ISCO/VSCO occupational mapping) before recommending LMIS-dependent mechanisms. See references/domains/education.md Section 5 for Vietnam-specific nodes.
Backward-looking system — mandatory check (REFINE): For any quota or allocation policy, explicitly map every criterion as backward-looking (past data) or forward-looking (projected demand). If 100% backward-looking, this is a structural gap that MUST appear in the analysis. Lag time from data collection to graduate output must be calculated (typically 5–6 years for 4-year degrees). Propose at least one forward-looking alternative from: ILO Skills Foresight, UNESCO Scenario Planning, World Bank Rolling Adjustment, or Pilot-Based Expansion.

NEW PITFALL: DO NOT conflate "Pilot-Based Expansion" with "low institutional cost." As identified in Vietnam policy analysis (TT34/2026), institutional cost is HIGH (not low) because pilot mechanisms typically conflict with existing legal frameworks (e.g., Law on Higher Education regarding institutional autonomy, Decree 60 regarding financial autonomy) and require high-level legal intervention or substantial "seed funding" from the state budget (e.g., South Korea's PRIME project). If proposing a pilot, you MUST propose the mechanism for legal/budgetary bypass (e.g., PM decision, budget reallocation). - Supply vs demand ratio — mandatory framing: Any enrollment quota analysis must answer: what percentage of the quota formula is supply-side (institution capacity) vs demand-side (labor market signal)? A system that is ~95% supply-side and ~5% demand-side has a structural demand-blindness problem regardless of individual proxy quality. - Design Failure vs Execution Noise — classify before recommending solution: - Design Failure (flaw in the text itself — missing definition, missing transition period): solution = amend the regulation. - Execution Noise (text is sound but guidance is lacking): solution = supplementary guidance. - Misclassifying Design Failure as Execution Noise leads to under-powered recommendations. - Proxy language: Do not say a proxy indicator is "meaningless" or "ineffective" unless it has zero operative value. The correct framing: "lacks independent verification mechanism to achieve full effect." This matters for advocacy — calling a proxy worthless alienates the officials who designed it and closes dialogue. - Source verification for legal citations: Always cross-check resolution/decree numbers against actual content. A number alone is insufficient — NQ36 can exist in multiple years with different subjects (e.g. NQ36/2023 = maritime economy, not education reform; the education reform NQ is NQ29/2013). Verify year + subject before citing. - NLM auth expiry — fallback immediately: When nlm notebook query returns "Authentication expired", do NOT attempt to re-authenticate interactively (it opens a browser that cannot be driven in a non-GUI session). Fall back to web_search and mcp__paper_search_mcp__search_* tools immediately. The NLM notebook still contains the best primary sources for Vietnam education policy — note which queries were skipped and flag them as [NLM query not executed — auth expired; verify from notebook e19cb17a] in the output so a follow-up session can fill the gap. - Multi-phase research inheritance: When producing a Giai đoạn 2+ analysis that builds on a prior Giai đoạn 1 document, READ the prior document in full before writing a single word. Explicitly track which findings are inherited (cite section), which are extended (new evidence), and which are new territory. Never repeat analysis already present in the prior phase — synthesize forward. - ASEAN benchmarking — country set and depth order: When benchmarking Vietnam education policy against ASEAN/Asia-Pacific peers, the canonical priority order by contextual similarity is: (1) Malaysia — dual public/private system, accreditation-led quality control; (2) Thailand — manpower planning integration, cooperative quota for FDI; (3) Philippines — moratorium on over-supplied programs, demand-pull scholarship steering; (4) Indonesia — large fragmented system, PDDikti data registry; (5) South Korea — financial incentive restructuring (PRIME), demographic-cliff-driven reform; (6) China — centralized planning, strategic sector targeting; (7) Singapore — demand-side anticipation model, skills-framework-driven intake. See references/domains/asean-education-benchmarks.md for detailed country profiles.

9. Domain Reference Files

Support files are in references/domains/. Load the relevant file(s) at task start — do not wait until mid-analysis.

File	Contents
`references/domains/vn.md`	Vietnam institutional context: 5 nút thắt, bẫy Thông tư, 6-step VĐCS workflow
`references/domains/education.md`	Education policy domain: OECD 3-pillar framework, key trends, data infrastructure, Vietnam-specific nodes (TT34, kiểm định, BHXH lag)
`references/domains/education-evaluation.md`	OECD WP236 evaluation framework: Theory of Change, 6-dimension grid, 5-step workflow, TT34/2026 application table
`references/domains/asean-education-benchmarks.md`	ASEAN/Asia-Pacific higher education enrollment benchmarks: 7 country profiles (Malaysia, Thailand, Philippines, Indonesia, South Korea, China, Singapore), supply/demand ratio estimates, and future structural mismatch risk forecast
`references/domains/vietnam-education-policy.md`	Vietnam enrollment quota domain: TT34/2026 mechanism, TT03 vs TT34 comparison, supply/demand ratio analysis, ASEAN LMIS context, backward-looking lag, forward-looking gaps

To add a new domain: create references/domains/<sector>.md and add a row to the Domain Directory table (Section 1). Do not modify core SKILL.md logic sections.