Take in a piece of information, pass through the surface to find the motif it repeatedly enacts, then draw clearly the few structures that are actually doing the work and send them into the wind tunnel for calibration. The default result should be light: the reader comes away with a motif that can be stated clearly, a relation diagram that can run, and a conclusion that both transfers and changes its mind.
Four concepts that must not be conflated
- Theme: what object the information is about, such as 'team discussion' or 'AI attention'. It only delimits the semantic scope.
- Motif: the same kind of fundamental predicament repeatedly encountered across different domains. It is not replacing the original topic with a broader abstract term, but preserving that transferable relation in more ordinary words.
- Analogy: two cases look alike, such as 'a team is like a jury'. It is only an entry point for discovery.
- Structure: the causal skeleton that still holds after the domain nouns are deleted. It explains what relation generates what result.
In one sentence: theme is the name, motif is the predicament, analogy is the clue, structure is the causal machine.
Understanding contract
Structural analysis is not about getting the reader to memorize a set of abstract nouns, but about enabling them to run the relations through in their mind. A completed interpretation must achieve:
- Clear concepts: every load-bearing concept can be explained in plain language as 'what it specifically refers to here', with a visible marker.
- Clear relations: the reader can restate 'under what conditions, who affects whom how, what change appears first, and what result is finally obtained'.
- Examples carry the relations: examples do not merely make people feel a similarity; the roles, directions of effect, and results within them all correspond to the abstract relations.
- Boundaries can be judged: when conditions change, the reader can predict how the relations will change and also knows what observations would force them to revise their judgment.
Cases are responsible for making the model run, not for proving the model universally true; the wind tunnel is responsible for changing conditions internally, finding boundaries, and calibrating judgment, and ultimately hands only the informative stress-test results to the conclusion.
How the wind tunnel enters the conclusion
- The wind tunnel is an internal method: continue using the same example, change one key condition, observe how the prediction changes, and find out when the model fails and what signal would make the judgment change its mind. It is no longer listed as a visible section.
- The conclusion is the calibrated deliverable: first present the key differences in the same example after the condition change, giving the reader something concrete to hold onto; then compress the differences into one transferable abstract rule, returning to the original phenomenon to explain what should now be seen.
The wind tunnel's full reasoning stays internal; the conclusion only brings out the condition changes, boundaries, or change-of-mind signals that best distinguish the model. They and the abstract rule must form a progression, rather than stating the same causal chain twice.
Applicable boundaries
Suitable for: unclear problem mechanisms, a lack of ready-made solutions, multiple phenomena affecting one another, or explorations that need to go beyond ordinary AB testing.
Not suitable for: purely factual queries, execution tasks with only one definite step, or inputs with no identifiable dilemma. In high-risk domains such as medicine, law, and safety, structure may be used to generate hypotheses, but analogy must not replace domain evidence.
Workflow Routing
| Workflow | Trigger | File |
|---|---|---|
| FindStructure | Find motifs, structural relationships, and wind-tunnel experiments from the information | Workflows/FindStructure.md |
Gotchas
- "Team communication" and "AI cognition" are still topics; "how to do better" is again so broad that it has no constraints. A motif must preserve the real relationship or tension.
- A motif is not a more abstract title. If a one-sentence motif still contains two or more load-bearing words that cannot be explained in plain language, decode them first, then output.
- Under the motif, place only one minimal example, used to let the relationship run for the first time; do not expand into a cross-domain case list.
- By default, set only one example anchor that runs through the entire text. Structure cards point back to specific moments within it, the relationship diagram connects the whole, and the wind tunnel changes its conditions; do not retell the same story in four places.
- "Like the immune system" and "adopt double-blind" are only sources or names of solutions. Only when the causality can still be stated clearly after removing the source name does it count as structure.
- Cross-domain search is internal skill, not a default table of contents. Laying out three domains item by item often only suppresses the real motif.
- By default, keep only one or two structure cards; a third must prove that it cannot be merged into the first two. Each card must complete "clear concept, clear relationship, example mapping," but it is not required to display them as three fields.
- Examples must map roles and causality. Saying only "this is like so-and-so" creates familiarity but does not make the relationship clear.
- When there are multiple structures, first determine whether they are serial, parallel, nested, balancing, or feedback, then draw the diagram; do not line up multiple structures into a chain just because you see them.
- By default, do not take a Cartesian product or calculate the number of combinations. With multiple motifs, first find the structural relationships; expand only when the user explicitly requests combinations.
- The wind tunnel is a method, not a chapter. It first locks in a baseline prediction, then changes one condition, compares the earliest difference to appear, and finally points out the applicable boundary and the rule for updating judgment; in the written text, keep only the one stress test that most changes understanding.
- The conclusion is not a summary of the whole text. It completes, in order, "specific stress test -> abstract rule -> return to the phenomenon," letting the example provide a handle and the abstract relationship gain transferability.
- The specific stress test and the abstract rule must progress: the former answers "what happens when conditions change," while the latter answers "what more general relationship this shows." If the two paragraphs merely restate the same thing with different words, delete the weaker paragraph.
- The minimal reversible experiment is used only for action problems. Conceptual interpretation by default gives a discriminating test, and does not turn every thought problem into an action recommendation.
- A conceptual model cannot evade counterexamples by narrowing its boundary after the fact. First declare the applicable conditions and distinctive predictions; cases outside the boundary only serve to draw the boundary and do not count as support or refutation.
- The wind tunnel does not introduce a new set of terminology. If new concepts must be invented to explain the stress-test result, first return to the structure card and decode again.
- The evidence boundary is stated only once at the end. Do not repeatedly insert warnings into every structure card, interrupting the reasoning.
- When facts are uncertain, keep them in the internal candidate layer; if they must be output, clearly write them as pending verification, and do not treat remembered stories as facts.
Output requirements
By default, about 800-1500 Chinese characters. The first-level heading is strictly only:输入、母题、结构、结论. A wind tunnel is an internal reasoning step and is not output as a first-level heading. The motif uses one question and one example so that the relationship starts running first; the structure card completes "clear concept, clear relationship, example mapping" in natural short paragraphs, without forcing visible fields. Cross-domain sources do not get a default heading; usually they are not output, and when they are truly helpful, use only one sentence with one example. When there is more than one structure, draw the relationship diagram with pure ASCII connectors; Unicode arrows and box-drawing characters are forbidden. After the relationship diagram, use one plain sentence to explain "how this diagram runs in the example."
By default, obtain two time values:date +%Y%m%dT%H%M%S and date "+%Y-%m-%d %a %H:%M". Write:
~/Context/{时间戳}--结构-{主题}__structure.org
If the user explicitly says "analyze only," "do not save to disk," or read-only, directly provide an equally complete analysis in the conversation and do not create a file.
The file must be org-mode; Markdown is forbidden. Minimal structure:
#+title: 结构：{主题}
#+date: [{可读时间}]
#+identifier: {时间戳}
#+filetags: :structure:
* 输入
* 母题
* 结构
* 结论
The conclusion closes in two natural paragraphs. The first paragraph returns to the same example anchor and changes only one key condition, writing out the first difference to appear and what result would force the model to revise its claim. The second paragraph extracts from that difference one transferable relationship, directly answers the motif, and then points back to the original phenomenon to explain how it should now be understood. At the end, the evidence boundary appears only once, distinguishing illustrative cases, structural derivations, existing support in the field, and parts still awaiting experiment. After saving, read the file back and report the path.
Examples
Example 1: The first speaker locks in the team
User: 「为什么团队讨论总被第一个人带偏？找其中的结构。」
→ 母题用普通话追问：为什么越早出现的信息越容易成为后续判断的参照物？
→ 用会议现场的一句话，让这条关系先运行一次
→ 结构卡说明「参照物」指什么、怎样影响后来者，并给具体例子
→ 内部风洞改变发言顺序或信息来源，比较预测怎样变化
→ 结论先给出条件变化后的具体差异，再抽出「先出现的信息会取得解释权」
Example 2: AI deep dive and anomaly discovery
User: 「这个 AI 擅长深挖局部，却总错过改变全局的异常，用结构风洞分析。」
→ 母题：有限注意力如何兼顾局部开发与未知探索
→ 抽出局部深挖与外围巡逻的关系骨架
→ 内部用风洞找出最小异常升级实验，把试压结果并入结论
Example 3: The user explicitly asks for expansion
User: 「请给出这个母题的跨域来源，并分析三个母题怎样组合。」
→ 才展开来源或多母题处理
→ 仍先判断结构间的串联、并联、嵌套、制衡或反馈并画图
→ 只有用户继续要求组合数学时，才计算组合空间