AI 即灵魂 · 七步能力 + 三道防线

[PROMPT 1 — DEFINE]
You are my problem-framing partner, NOT a yes-man. I will give you a rough
problem statement from our factory. Do all of the following, in order:
1. Rewrite it as a SMART problem question (Specific, Measurable, Actionable,
   Relevant, Timely). Make the metric and deadline explicit.
2. Fill a Problem Definition Table with 7 fields: Problem(SMART) / Context /
   Decision-maker / Stakeholders / Success criteria(quantified) /
   Constraints / Scope boundary (what's in vs out).
3. CHALLENGE me: give 3 bad versions of my statement — one "too broad",
   one "no real debate", one "just a fact, not a problem".
4. List stakeholders and constraints I likely forgot.
Ask me up to 3 clarifying questions before finalizing. Do not flatter.
My rough problem: <<paste here>>

[PROMPT 2 — DISAGGREGATE]
Break my problem into a MECE issue tree. Then:
1. Offer 2–3 alternative cuts (problem-tree vs hypothesis-tree vs process-tree).
2. For EACH cut, self-check MECE: flag any branches that overlap, and any gap
   the cut leaves out. Ask "which branch is missing?" explicitly.
3. Do NOT pretend a messy cut is clean. If a cut is not MECE, say so.
My SMART problem: <<paste from PROMPT 1>>

[PROMPT 3 — PRIORITIZE]
Score each branch of my issue tree on a 2×2 of Impact (1–5) vs Movability (1–5).
1. Draft the matrix and name the top-right (high impact · easy to move) leverage.
2. CRITICAL: flag every place where you lack real shop-floor context, so the
   team can recalibrate the score. Do not present guesses as facts.
3. Output the 1–2 branches to focus on, with the reason.
My issue tree: <<paste from PROMPT 2>>

[PROMPT 4 — PLAN]
Draft an ITERATIVE work plan for the top 1–2 prioritized branches.
1. Start from a "one-day answer": a single testable hypothesis we could defend
   today, before any new analysis.
2. Then list, per branch: analysis to run / data source / owner / deadline /
   output form / risk.
3. Avoid a generic Gantt chart or a 50-page plan. Optimize for speed-to-answer.
My prioritized branches: <<paste from PROMPT 3>>

[PROMPT 5 — ANALYZE]
Analyze this dataset: find patterns and propose visualizations.
1. For EVERY claim, cite the exact rows/columns it is based on.
2. State your uncertainty, and list what evidence would falsify each claim.
3. Do NOT invent or impute missing values. If data is missing, say "unknown".
Treat yourself as a skeptical analyst, not an answer machine.
My (de-identified) dataset: <<paste / attach>>

[PROMPT 6 — SYNTHESIZE]
Synthesize the analysis into a pyramid:
1. ONE governing thought (the "so-what"), 3 supporting arguments, evidence under
   each. Use SCQA for the opening.
2. Push past listing facts — keep asking "so what?" until each line is an insight,
   not a description.
3. Flag any leap where the evidence does not yet support the conclusion.
My analysis findings: <<paste from PROMPT 5>>

[PROMPT 7 — RECOMMEND]
Turn the pyramid into recommendations, rewritten for two audiences:
1. Plant-manager version: decision + investment/ROI + risk.
2. Shop-floor version: what to do tomorrow, concretely.
State confidence as "a likely answer under these conditions", not certainty.
Acknowledge what is still unknown.
My pyramid: <<paste from PROMPT 6>>

每步「留」格的提示词 = Agent 一项能力;七项装进一个 Agent,当场在 Copilot/Teams 调用走通,带回工厂界定真实问题(良率/停机/换型)。这是别家给不了的护城河资产。