surface-insight（表面洞察力）

Surface Insight

Take data, observations, or notes and surface insights that were hiding in the data all along. The goal is not theory — it's the "obvious in hindsight" moment where a connection clicks into place because the evidence was right there, just unseen.

When to Use

• User provides bullet points, notes, data, or observations and wants to extract meaning
• User asks "what does this mean?" or "what's the insight here?" or "connect the dots"
• User has disparate information and wants to understand how pieces relate
• User wants to reason about implications, predictions, or hidden dynamics

Do NOT use for: summarizing (use sharpen-prompt), writing articles (use write-oped), or evaluating prompts (use think-critically).

The Grounding Rule

Every insight must be rooted in the user's actual data. This is the single most important rule:

• Every insight must cite or quote specific data points the user provided
• If you cannot point to concrete evidence in the input, the insight is theory — discard it
• The test: "Could someone verify this insight by re-reading the user's data?" If no, it fails
• Prefer insights where the evidence is *already present* but the connection is *not yet drawn*

An insight that is well-grounded but modest beats a spectacular insight with no evidentiary anchor.

Process

Phase 1: Intake (Silent)

Read the user's data. Silently identify:

• What's surprising, anomalous, or unexplained in the data
• What data points seem unrelated but might not be
• What's conspicuously absent given what is present
• Implicit assumptions or framing the user may have

Do not output Phase 1.

Phase 2: Data-First Discovery (Silent)

Do NOT start from the lenses. Start from the data:

1. Find the 3-7 most interesting tensions, anomalies, or unexplained connections in the data
2. For each, ask: "What would make this click into place?"
3. Only then check the Reasoning Lenses below to find vocabulary and structure for what you already found

The lenses are a toolkit for articulating insights, not a checklist for generating them. If a lens doesn't illuminate something already in the data, skip it.

Do not output Phase 2.

Phase 3: Insight Development

For each discovery, develop the insight:

1. Ground it — Quote or cite the specific data points you're connecting
2. Name the connection — State the non-obvious relationship clearly
3. Show why it clicks — Walk through the reasoning so the reader can verify it against the data
4. Assess confidence — Speculative, plausible, or well-supported

HARD RULE: Output exactly 3-7 insights. If you find yourself generating an 8th, discard the weakest. Three strong insights beat seven mediocre ones.

QUALITY GATE: Before presenting each insight, ask: "If I showed this to the user, would they re-read their own data and say 'how did I miss that?'" If no, discard it.

Phase 4: Output

Present insights ordered by click-factor (strongest "obvious in hindsight" first).

Output format:

### [Insight Title]

**From the data:** [quote or cite the specific data points being connected]

[2-4 sentences: the insight and reasoning chain. Every sentence must reference something concrete — a named entity, a number, a mechanism, a specific outcome. No sentence may consist entirely of abstract generalities.]

**Confidence:** [speculative | plausible | well-supported]

End with a ## Synthesis section only if the individual insights compound into a higher-order conclusion none of them states alone. If no such pattern exists, omit entirely rather than forcing one.

ANTI-PATTERN CHECK: Before presenting, scan each insight against the Anti-Patterns below. Revise or discard any that fail.

Reasoning Lenses

Use these as vocabulary for insights you've already found in the data — not as a generation checklist.

Connect — Find hidden links

• Convergence. Independent developments heading toward the same collision point.
• Cross-Impact. A affects B through a mediating mechanism not visible on the surface.
• Triangulation. Multiple independent signals pointing to the same conclusion.

Project — Extend data through time

• Extrapolation. Current trends projected forward to a concrete prediction.
• Rate-of-Change. The velocity or acceleration of a metric is itself the signal (e.g., still growing but decelerating).
• Regime Change. The system has shifted qualitatively, not just quantitatively — old rules no longer apply.
• Temporal Displacement. What's newly possible (or impossible) because conditions changed?

Explain — Build and test causal stories

• Abductive Reasoning. If A and B are both true, the best explanation is H — and here's evidence for H.
• Retroduction. If thesis T is correct, we'd expect evidence E1, E2, E3 — check which exist in the data.
• Counterfactual. Remove X from the picture — does the outcome still hold? If yes, X wasn't causal.

Reframe — See from a different angle

• Analogy. Structural parallel to another domain or era, with transferable lessons.
• Disanalogy. The popular analogy breaks down — and the breakdown itself is the insight.
• Inversion. Flip the question to reveal blind spots.
• Dialectical Synthesis. The dominant narrative and its contradiction resolve into a higher-order truth.

Reveal — Expose hidden structures

• Second-Order Effects. The first-order impact is obvious; the second-order consequence is not.
• Incentive Mapping. Puzzling behavior becomes rational when you see who benefits.
• The Dog That Didn't Bark. The absence of an expected signal is itself a signal.
• Constraint Identification. The binding bottleneck that limits the entire system.

Insight Quality

The quality bar is the click test: does the insight make the reader re-examine the data and say "of course — how did I miss that?"

• Fail — Restatement. Rephrasing a data point. REJECT.
• Fail — Obvious. Something anyone would see. REJECT.
• Pass — Non-obvious connection. Linking data points that aren't obviously related. MINIMUM.
• Good — Hidden dynamic. Revealing a mechanism or feedback loop not visible on the surface.
• Strong — Assumption inversion. Overturning a default assumption with evidence from the data.
• Exceptional — Structural truth. Reframing the entire dataset.

Aim for "Good" or above on at least half of insights. Never output "Fail" level.

Anti-Patterns

• Theory without evidence. An insight that sounds smart but doesn't point to specific data. The #1 failure mode.
• Forced connections. Straining to link genuinely unrelated data points. "No strong insight here" is a valid output.
• Vague gesturing. "This could have major implications" without saying what they are.
• Prediction without mechanism. "X will happen" without showing the causal chain from the data.
• Pattern over-fitting. Seeing patterns because you're motivated to find them. Test: "Would I believe this if I hadn't been looking for it?"
• Hedging into uselessness. "This might possibly suggest..." — commit to the insight or drop it. Use the confidence label instead of hedging the language.
• Kitchen-sink analysis. Using too many lenses. 3-7 insights, not 18.

Key Principles

1. Data first, lenses second. Find what's interesting in the data, then use lenses to articulate it. Never the reverse.
2. Obvious in hindsight. The best insight makes the reader say "how did I miss that?" because the evidence was right there.
3. Every insight must point to evidence. If you can't cite specific data from the input, it's theory, not insight.
4. Concrete and specific. Not "this market could grow" but "this creates a $X opportunity in Y segment because Z."
5. Less is more. Three grounded insights beat seven theoretical ones.
6. Connect, don't summarize. The input is data. The output is meaning. The gap between them is where your work happens.

Input

[User provides data, observations, bullet points, or notes below]