feasibility-decomposition

Imported from zuoa/aj-skills (skills/feasibility-decomposition/SKILL.md). Install upstream with npx skills add zuoa/aj-skills --skill feasibility-decomposition. Copyright stays with the author.

name: feasibility-decomposition description: Deep feasibility analysis and structured decomposition of a concept/proposal/设想 against a reference framework (a policy, standard, plan, or spec). Produces (1) an alignment mapping matrix that surfaces coverage gaps, (2) a multi-dimensional feasibility assessment (technical / organizational / compliance / economic) with ratings and the domain's killer constraints, (3) value/highlights framed for award or funding applications, (4) a methodology-driven decomposition (default: the V字模型 with 业务协同模型 + 数据共享模型 double-through), and (5) standardized delivery checklists (business functions, data dependency & sharing, infrastructure, reform/transformation items). Use this whenever the user wants to 可行性分析 a 课题/设想/方案/产品 against a 政策框架/标准/规划/行业规范, do 数字化改革拆解, produce 业务功能清单/数据依赖及共享清单/基础设施依赖清单/改革变革项清单 for 课题申报/奖项申报/项目立项/可研/初设, or apply the V字模型/业务协同模型/数据共享模型. Trigger even when the user does not name the methodology — any request like "深度分析这个设想的可行性并拆解成清单", "对标 XX 框架/政策做拆解", "把这个课题按数字化改革拆解", or "帮我准备申报材料" qualifies. Also use for analogous tasks outside digital reform: any "concept measured against a framework, then feasibility + structured breakdown + checklists" request.

Feasibility Analysis & Structured Decomposition

You are given a concept (设想 / 课题方案 / 产品构想 / 方案) and a reference framework (政策 / 行动计划 / 标准 / 规划 / 规范 / 技术规格) to align it against. Your job is to turn the two into a rigorous, decision-grade deliverable: is this concept feasible, how does it map to the framework, where are the gaps, what's award-worthy, and what exactly must be built/reformed — broken down into concrete checklists.

This is a thinking-and-writing skill. The quality bar is high: outputs are typically used for 课题申报 / 奖项申报 / 立项可研, where vague analysis fails and where a single missed compliance red-line or coverage gap can sink the whole application.

The one-line method

设想 × 框架 → 对齐 → 可行性 → 亮点 → 拆解 → 清单

Walk the seven phases below. Each phase exists for a specific reason — don't skip, because later phases depend on earlier findings (gaps feed feasibility, feasibility feeds highlights, both feed decomposition).

Before you start: ingest and verify (do not hallucinate the framework)

1. Read the concept. If it's a .docx, extract text (a .docx is a zip; word/document.xml holds the text — a small Python zipfile + xml.etree snippet works, or use the docx skill). If .pdf, read it. If described inline, use that. Restate it crisply in your own words so you've actually understood it.
2. Fetch and verify the framework. This matters more than people think. Do not reconstruct a policy/standard from memory — fetch the real source. If a URL is given, read it (web reader / fetch). If the direct fetch is filtered or fails, fall back to web search to recover the framework's actual structure and requirements, and cite the source. Misquoting a national policy in a 申报 material is an unrecoverable error. When you cite a requirement, mark whether you verified it against the source (cite 文号 / 条款号 / URL) or are assuming it from the user's summary — don't blur the two. A reviewer who checks a phantom clause is a fatal error, so keep what's nailed-down visibly distinct from what's provisional.
3. Bound the scope. If the concept or framework is ambiguous (e.g., "is this for the whole city or one school?"), ask one focused clarifying question. Don't rabbit-hole on many questions — most can be resolved with a sensible default that you state explicitly and move on.

Phase 1 — Alignment mapping (对标对齐)

Build an alignment matrix: the framework's requirements/dimensions down the rows, the concept's coverage across the columns, each cell scored for coverage (e.g., ★★★★★ strong / ★★★☆☆ partial / ☆☆☆☆☆ gap), with a final column for gap + recommendation.

Why this comes first: the alignment matrix is where coverage gaps surface, and gaps are the most common reason an application loses points or a project fails review. Finding "the framework requires X but the concept doesn't address it at all" early lets you either close it (recommend adding it) or flag it as a known limitation — both are far better than the reviewer finding it for you.

Always end Phase 1 with 2–3 headline conclusions (strategic fit level, the must-close gaps, the standout alignment that becomes a highlight).

Phase 2 — Feasibility, deep and multi-dimensional

Assess feasibility across at least four dimensions, each with an explicit rating and a one-line judgment. Read references/feasibility-rubric.md for the full rubric, the rating scale, and how to find the killer constraint of the domain.

The dimensions (adapt/extend to the domain):

• Technical — by architecture layer or component; what's mature vs. needs R&D.
• Organizational / operational — who must cooperate, is that cooperation sustainable, where's the human bottleneck.
• Compliance & ethics — the most under-weighted and most fatal dimension. For minors + health data, for finance, for safety-critical systems, this is where projects die. Name the specific laws/standards and the specific red-line.
• Economic / sustainability — cost vs. budget, cold-start problem, who pays ongoing.

The single most valuable habit in this phase: find the domain's killer constraint — the one thing that, if unresolved, makes everything else moot (e.g., AI output crossing into 诊断/诊疗 triggers medical-device regulation; a self-improving "flywheel" can't cold-start without enough data). State it plainly and give the de-risking move. When the killer constraint is a binary design choice ("does this feature cross a regulated line?"), present the de-risking as a structured either/or: Path A (stay outside the constraint — what you give up) vs Path B (accept the constraint — what it costs in time/money/scope), with a recommendation. A forced decision is far more useful to the reader than scattered mitigation hints.

End with a feasibility verdict table (per-module rating + one-line定性).

Phase 3 — Highlights / value framing (亮点提炼)

Frame the concept's strengths in the language evaluators actually use. The reliable four lenses:

• Policy fit (政策契合) — how it advances the framework's goals, especially any source-stage / upstream positioning others miss.
• Model innovation (模式创新) — what structural collaboration or closed-loop is novel vs. the framework's own mechanism.
• Technical advancement (技术先进性) — the hardest-to-copy capability.
• Measurable outcomes (成效可量化) — anchor to the framework's own quantitative targets where possible (e.g., its 2030 KPIs).

Plus two that consistently differentiate winners: replicability / standardization (can others copy this path?) and ethics/compliance leadership (did they solve the hard governance problem others ignore?).

For each highlight, give the one-line claim AND the concrete evidence that would prove it (an indicator, a number, a artifact). A highlight with no proof is filler. At least one highlight must be anchored to a quantitative target — ideally the framework's own KPI (its 2030 target, a coverage rate, a decline percentage, a budget threshold). "Aligns with national policy" with no number attached is not a highlight.

Phase 4 — Methodology-driven decomposition (结构化拆解)

Apply a structured decomposition that threads two models through every part of the work. The default methodology is the V字模型 (浙江数字化改革): self-top-down decomposition + self-bottom-up integration, with the 业务协同模型 (business collaboration) and 数据共享模型 (data sharing) as the twin spines running through the whole V. Read references/decomposition-and-checklists.md for the V字模型 step-by-step and the adaptation rules.

Why two models, not one: business collaboration answers "who does what, when, reformed how"; data sharing answers "what data, from whom, shared how, governed how." In reform projects these two are inseparable — a beautiful workflow with no data pathway is fiction, and a data lake with no reformed workflow is a graveyard. Threading both is what separates 数字化改革 from mere 信息化, and it's where the "reform gold content" (改革含金量) lives that award panels look for.

Adapt when the task is NOT digital reform. If the framework is an industry standard, a technical spec, a compliance regime, or a research methodology, do not force the V字模型. Instead use that domain's native decomposition (e.g., a spec's clause structure, a standard's requirement categories, a methodology's stages). The constant is: decompose the concept into buildable/verifiable units along two complementary axes (one process/workflow axis + one data/information axis), threaded through every phase. See the adaptation box in the reference.

Phase 5 — Standardized delivery checklists (标准化清单)

Produce the checklists the decomposition implies. The default set (digital reform / 课题申报 context) is four lists:

1. 业务功能清单 (business functions) — id, name, layer, description, role, trigger, related data.
2. 数据依赖及共享清单 (data dependency & sharing) — entity, items, source, collection method, sharing parties & purpose, frequency, sensitivity, compliance.
3. 基础设施依赖清单 (infrastructure) — by layer (sensing / network / compute-storage / platform-AI / integration / portal / security).
4. 改革变革项清单 (reform & transformation) — by category (制度重塑 / 流程再造 / 规则重构 / 机制创新 / 标准规范), each with what it breaks from the old mechanism.

Each list is a markdown table. Full column schemas and adaptation rules (how to change the four categories when the framework isn't digital reform) are in references/decomposition-and-checklists.md, and the complete document skeleton including every table is in references/output-template.md.

Why four lists and not one: they map cleanly to the questions every reviewer/PM asks — what does it do (functions), what data does it need and from whom (data), what must we buy/build (infrastructure), and what rules/must we change to make it real (reform). A project that lists features but no reform items is IT, not reform.

Phase 6 — Risks & phasing (optional but recommended)

Close with a one-page risk countermeasure table (risk / level / mitigation) and a phased rollout (phases with period, focus, exit goal). Phasing especially matters when the concept has self-improving/learning components that cannot cold-start — sequence them behind a manual/half-auto phase until data volume crosses the threshold.

Output

Produce one markdown file named descriptively (e.g., <concept>_可行性分析与拆解.md), placed in the working directory, following references/output-template.md. After writing, give the user a tight summary of the headline findings (the gaps, the killer constraint, the top highlights, the list counts) and offer 2–3 concrete next steps (convert to Word/PDF, deepen one list into a 可研/初设 table, design the indicator system). Offer; don't auto-do.

Quality bar — what separates a great run from a mediocre one

• The framework is real, not invented. Cite it. If you couldn't fully verify a requirement, say so.
• Gaps are named, not hidden. A gap you surface is a problem you're helping solve; a gap the reviewer finds is a failure.
• The killer constraint is identified. Every domain has one. Find it.
• Feasibility is rated, not vibes. Use the scale. A reader should see at a glance what's green, what's yellow, what's red.
• Reform items exist when the context is reform. No 改革变革项 = you did IT planning, not reform planning.
• Highlights have proof. Claim + evidence, always paired.
• Two models threaded, not bolted on. Business collaboration and data sharing appear in every phase, not just one section.

When to deviate

If the user only wants part of this (just feasibility, or just one list), do that part well rather than forcing the full seven phases — but say what you're omitting by enumerating the omitted components by name with a reason. For example: "omitted — V字模型 decomposition, the business-function / data / infrastructure lists, and the highlights section, because the classification path must be decided before function lists are meaningful." Naming ≥2 omitted components with a rationale is far more useful than "I'll keep it short" — it tells the reader exactly what's missing so they can ask for it.