Agent Skills: Capture Note (Insight / Principle / Hypothesis / Question / Thought Mode / Idea)

Capture Note (Insight / Principle / Hypothesis / Question / Thought Mode / Idea)

These six note types share one shape — a template + type-specific frontmatter, filed in pages/ of whichever vault the content belongs to — but they answer different questions and their biggest value is in not confusing them with each other. This skill always classifies first, even when the user already named a type: the user's label is a hypothesis about the content, not a given.

Meta notes (all in the main vault /Users/zhaoliang/LocalDocuments/vaults/vault/pages/): insight meta.md, principle meta.md, Hypothesis Meta.md, questions meta.md, thought mode meta.md, ideas meta.md. Thought Mode notes are rare — the vault currently has the meta note but no instances yet, so lean conservative on this gate (see §4).

0. Classify — run this even if the user already said the type

Ask, in this order (each "yes" short-circuits — don't keep checking once one fits):

1. Is it a concrete, buildable thing — a project/product/tool/hack you could actually go make — and not a claim, judgment, or question at all? → Idea. The test: can you say "the thing I'd build/do is ___" in one sentence? If it's still a fuzzy direction ("do something with AI"), it isn't concrete enough yet — ask for the concrete version before filing, or don't file it.
2. Is there no claim yet — just an unresolved uncertainty that needs time, not information, to answer? → Question. (If it can be resolved by looking something up, it's not a Question at all — just answer it.)
3. Is this the same realization/mistake recurring, or a hypothesis that's gone long unrefuted, and it's already quietly steering choices? → Principle candidate — but only if a real boundary exists (step 4 of the principle section). If recurrence can't be shown, it isn't ready to be a Principle yet; fall through to whichever of insight/hypothesis fits and say so.
4. Is it a dated belief-shift about yourself — "I used to think X, now Y" — that already happened? → Insight.
5. Is it a forward-looking or world-level claim that could turn out wrong — a prediction, a judgment about someone/something external, something you're fairly confident of but haven't locked in? → Hypothesis.
6. Is it not a conclusion at all, but a reusable way of approaching problems — something that would apply the same way across unrelated domains (assumption-testing, Bayesian updating, steelmanning), that shapes how judgments get formed rather than being a judgment itself? → Thought Mode. This is the rarest of the six — most things that feel like a "mode" are actually a Principle (decision stance) or a one-off technique that doesn't clear the bar (see §4's gate). Default to not creating one unless it clearly fails to fit anywhere else.

A concrete deadline-bound task ("I'm doing X next week") fails the Idea gate too — that's a to-do, not a someday/maybe spark; point the user at their task manager instead of filing it here.

0.5 Choose the vault — run this after classifying, before writing anything

Sessions increasingly run from all-vaults, a view-only container that symlinks every real vault in (see its own CLAUDE.md) — it holds no content of its own, so never write a new note directly under all-vaults/. Always resolve to the real vault directory the content belongs in and write there (that's also where any git commit for it happens).

The meta notes always stay in the main vault (vault/pages/*.md) — they're the one authoritative definition each type has. But individual captured instances don't default to the main vault; which vault an instance belongs in is content-dependent, and gets decided fresh each time:

1. Read all-vaults/CLAUDE.md's vault table to shortlist candidates by domain (career/job-search → job-hunting-vault, money/business → wealth-vault, AI research → AI-vault, networking/CS research → survey-vault, sexuality/intimacy → entertain-vault, books → book-vault, language learning → language-vault; general self-reflection/life content with no sharper home → the main vault).
2. If a candidate vault has its own CLAUDE.md (self-analysis-vault, survey-vault, job-hunting-vault currently do), read it before deciding — it can define a more specific ownership rule than the domain guess (e.g. self-analysis-vault's 认识论网格) that should win.
3. If nothing clearly fits a domain-specific vault, default to the main vault — don't force a note into a niche vault on a weak match.
4. State the chosen vault in the report-back, and why, whenever it isn't the main vault — this routing call is worth surfacing, not just silently acting on.

Broken cross-vault wikilinks are an accepted tradeoff, not a reason to keep content in the wrong vault. all-vaults gives cross-vault search, so a link that won't resolve inside one vault's own graph is findable another way — it doesn't justify leaving a self-analysis note sitting in the main vault (or vice versa) just to keep a [[...]] clickable. Route by topic first; let links break.

This applies to all six types, but matters most for Insight / Principle / Question / Hypothesis, since those are the ones most likely to be about a specific life domain (career, money, research) rather than general self-reflection.

If the user's stated type doesn't match what step 0 concludes, say so plainly before writing anything — name which criterion it fails and which type actually fits — and confirm before proceeding. Don't silently "correct" them, and don't force a note into a type just because they named it. If nothing fits any of the six (e.g. it's a plain fact, a how-to, a feeling, a slogan), say that too and suggest the right home (reference/tip/journal) instead of forcing a capture.

1. Insight

Implements insight meta.md: a real, dated belief-shift — not information, not a quote, not a feeling, not an unqualified claim about the world.

Gate: must have a real Before/After. If the user can't state what they used to believe, ask — don't invent a Before to force the shape.

Frontmatter:

---
created: YYYY/MM/DD, HH:mm:ss
last_reviewed: YYYY/MM/DD
modified: YYYY/MM/DD, HH:mm:ss
status: candidate
tags:
  - insight
title: <same as filename>
---

status: candidate always at capture time — it only moves (→ principle, → hypothesis, → dropped) on a later review, per the meta note's lifecycle.

Body (template/mind/New insight.md's structure — don't paraphrase into flowing prose):

# <title>

# What Changed

我之前以为：<the old belief, one line>

现在意识到：<the new understanding — what actually triggered seeing it, e.g. a specific device,
measurement, or moment, not just the abstract realization>

# Why This Matters

<what this changes about a real decision/behavior — concrete, not "this is important because...">

# What I'll Watch for

接下来我会观察：
- <a concrete situation where this belief will get tested or applied next>
- <a plausible counter-example or alternative explanation that would undermine it>

## 相关

- <wikilinks to related concept pages>
- 来源：<journal date + Quick Notes timestamp if this came out of a journal entry / conversation>

2. Principle

Implements principle meta.md: a decision stance validated more than once —

当我犹豫时，它能稳定地替我做选择。

Gate (the hardest one in this skill). Create only when at least one holds:

• an Insight has recurred (same realization/mistake, more than once)
• or a Hypothesis has gone long without being refuted
• and it's already influencing choices without being written down
• and doing the opposite would produce clear regret

Reject and redirect when: it happened once (→ Insight), it's a slogan/value statement with no decision content (要严谨 etc.), it's a prediction about the world (→ Hypothesis), it's a procedure (先lint再commit → tip/workflow), or it has no findable boundary (an "always do X" with no counter-case is a moral demand, not a principle).

Ask for recurrence evidence before writing anything: "这条以前在哪些场合出现过？" If the user can't name a second occasion, say so and offer Insight instead, promotable later. If promoting from an existing Insight note, read it first — its Before/After is the principle's origin, and its "What I'll Watch for" often already contains the boundaries.

Location: run §0.5 to pick the vault, then prefer the same vault as the Insight/concept notes it cites so wikilinks resolve — flag in the report-back if that pulls it away from what §0.5's domain call would otherwise pick.

Title = the decision, not the topic (差值不是判据除以SE才是, not 标准误).

Frontmatter:

---
created: YYYY/MM/DD, HH:mm:ss
modified: YYYY/MM/DD, HH:mm:ss
tags:
  - principle
title: <same as filename>
---

Body (template/mind/New principle.md's structure):

# <title>

> **<one-sentence decision statement: when I hesitate, what do I default to?>**

<optional: the formula / the concrete test, if the principle has one>

## Why this matters?

<the recurring bind it resolves, with the concrete cost of getting it wrong — a number or a real
outcome, not "it's important">

## When it applies

<trigger signals: the words or situations that should make this fire>

## When it does NOT apply

<real boundaries — mandatory, must be non-empty (see below)>

## When I violate this

<recurrence evidence — prefer a table>

| # | 场景 | 我说的 | 实际 |
|---|---|---|---|

## 相关

- <wikilink to the Insight/Hypothesis it was promoted from, and related concept pages>

When it does NOT apply is mandatory. Before writing, actively look for: the case where following it is technically correct but answers the wrong question; where its precondition fails; where it can't be executed (the needed input doesn't exist); where it degenerates (works but stops discriminating). If none can be found, don't create the Principle — capture as Insight instead.

Link back: pointer line on the source Insight (- ⚖️ 已升级为原则：[[<principle>]]（<why>）); one-line hook (⚖️ [[<principle>]] —— <hook>) on concept pages where it's needed at decision time, not everywhere the topic is mentioned. Update the source Insight's status: principle only if the user wants that.

3. Hypothesis

Implements Hypothesis Meta.md: a temporary judgment about reality —

我目前认为世界可能是这样，但我允许自己被证明是错的。

Gate: forming a judgment about the future/trend/other people/the world, that could matter to a real decision, held with more confidence than the evidence justifies, or compressing complex reality into one sentence. If it's a value stance about how to live → Principle, not Hypothesis. If it's already stably guiding decisions → Principle. Must have a falsification condition — if you can't write "if X happens, I was wrong," it's a belief, not a hypothesis; don't create it.

Frontmatter (Templater-style, fill placeholders directly since this isn't run through Obsidian):

---
tags:
  - hypothesis
status: 未采信
confidence: <0-100, a number — not L/M/H, numbers are the only thing you can audit against later>
importance: <1-5 decision leverage: would being wrong change a real choice? 5 = changes track/geography, 1 = don't bother filing>
created: YYYY-MM-DD
review:
---

Body (template/mind/New hypothesis.md's structure):

# {{title}}

> **status** 未采信 / 部分采信 / 已采信 / 已推翻
> **confidence** 数字，事后对账用
> **importance** 决策杠杆 1–5
> 取证优先级 ≈ importance × 不确定度（confidence 离 50 越近越该查）

## Hypothesis

我当前的判断是：

<one falsifiable sentence — if it doesn't fit in one sentence, it bundles several claims; split it>

## Source & Grade

| 来源 | T级 | 日期 | 该打什么折 |
| ---- | --- | ---- | ---------- |

<T3 for secondhand/retold sources until traced to the original>

## What Would Prove Me Wrong

<write this NOW, before more evidence comes in — specific and observable ("if X happens"), not
"if the situation changes". No falsification condition = belief, not hypothesis (recheck the gate).>

-

## Next Evidence to Collect

<ranked by cost/value — cheap-and-decisive first. Don't list "keep reading similar secondhand
articles" — that raises confidence-of-feeling, not actual confidence.>

- [ ]

## Decisions This Affects

<name the actual decision this feeds, e.g. which blueprint section. Can't name one = this is a
passing observation, not a hypothesis — don't file it.>

## Log

| 日期 | 发生了什么 | confidence |
| ---- | ---------- | ---------- |
| YYYY-MM-DD | 建档 | |

Most hypotheses should sit at status: 未采信/confidence candidate-level for a long time — resist the pull to mark things 已采信 early just because they feel obviously true.

4. Thought Mode

Implements thought mode meta.md: a reusable way of thinking, not a conclusion —

当我面对复杂、不确定或高风险问题时，我选择用"哪一种方式"来思考。

It's the "运行模式" of the cognitive operating system, not an opinion or an answer.

Gate — the rarest type, keep this tight. Reject and redirect when: it's a specific conclusion or judgment (→ Insight/Hypothesis), a value stance (→ Principle), a step-by-step procedure (→ tip/workflow), a one-off technique that doesn't generalize across domains, or an academic concept definition dressed up as a "mode." The test that actually discriminates: would this same mode apply, unchanged, to a completely unrelated domain (research methodology, a relationship conflict, a financial decision)? If it only makes sense inside the one context that produced it, it's not a Thought Mode yet — it's a technique, and techniques don't get filed here.

Since this vault has zero instances so far (only the meta note), don't lower the bar to populate it — a Thought Mode not yet used across ≥2 unrelated domains isn't ready; note it as a candidate-in-waiting in the report-back rather than filing it.

Frontmatter:

---
created: YYYY/MM/DD, HH:mm:ss
modified: YYYY/MM/DD, HH:mm:ss
status: candidate
tags:
  - thought-mode
title: <same as filename>
---

status: candidate at capture time — moves to active only once it's shown up unprompted across multiple scenarios (per the meta note's Status Semantics), never at filing time. Don't have too many active modes at once — the meta note warns this causes 思维瘫痪 (decision paralysis).

Body (Core Elements from thought mode meta.md — no Templater template exists for this type, so use this structure directly):

# <title>

## Core idea

<what orientation of thinking this is, in one or two sentences>

## When to activate

<what situation/trigger should make you switch into this mode>

## How to use

<a short mental checklist or concrete operating prompt — must be executable, not philosophical>

## Failure modes

<how this mode commonly gets misused — including "used to avoid a decision" if applicable>

## 相关

- <wikilinks to the Insight/Hypothesis/decision this mode has actually been applied to>
- 来源：<where this crystallized — journal date, conversation, reading>

5. Idea

Implements ideas meta.md: a concrete spark you could go build/try/make —

如果做了 X，会怎样？——我现在没空/没资源去做，但不想忘记它。

Unlike the other five types, an Idea isn't a claim, a stance, or a question at all — it's generative, not evaluative. It doesn't need confidence tracking, falsification conditions, or Before/After — its only job is to not get lost.

Gate: must be concrete enough to state as "the thing I'd build/do is ___" in one sentence. Reject and redirect when: it's a deadline-bound task you're actually doing (→ task manager, not a someday/maybe spark), it's a fuzzy direction with no concrete "what" yet (ask to tighten it first), it's a claim about how the world works (→ Hypothesis), or it's a belief-shift about yourself (→ Insight).

Frontmatter (Templater template exists — template/mind/New idea.md):

---
tags:
  - ideas
status: candidate
created: YYYY/MM/DD, HH:mm:ss
---

status: candidate at capture time; it only moves later (→ tried → shipped/merged/dropped) on review, per the meta note's Lifecycle — never mark it shipped at filing time.

Body (template/mind/New idea.md's structure):

# <title>

<the idea itself — one sentence stating concretely what you'd build/do; the title is usually
already this sentence, expand by a line or two only if needed>

## 为什么有意思

<optional at capture time — can be left blank. What problem does it solve, or what makes it worth
keeping?>

## 第一步大概怎么做

<optional. Not a full plan — just enough that the thread isn't lost: if you actually tried this,
what's the smallest first step?>

## 相关

<related links, inspiration sources; 来源：journal date + Quick Notes timestamp if it came out of a
journal entry / conversation>

Most Ideas should stay bare — title plus whatever fell out of the moment. Don't force the "为什么 有意思" / "第一步" sections to be filled in just to look complete; leaving them empty is normal and matches how this vault has actually been using the ideas tag so far.

6. Question

Implements questions meta.md: a deliberately open question —

我还没有答案，但这个问题值得我反复带着。

Gate: it needs time, not information — if it's answerable by a lookup, it's not a Question. Reject: things with a stable answer already, emotional venting disguised as a question, "fake" questions where the answer is already decided, questions used to manufacture anxiety.

⚠️ There is currently no Templater template for this type (New hypothesis/insight/principle.md exist in the shared template repo; New question.md does not) — use the structure below directly, matching the pattern used in pages/AI时代科研工作者应该如何适应.md.

Frontmatter:

---
aliases: [<the question, phrased as the user said it>]
created: YYYY/MM/DD, HH:mm:ss
tags:
  - questions
status: open
title: <the question, as a short imperative/interrogative title>
---

Body:

# <title>

## Clear question

<specific, understandable — split into separate Questions if it's actually bundling more than one>

## Why it matters

<why this is worth carrying long-term — link it to what it would actually change, not "it's
interesting">

## Context

<what stage/situation it surfaced in>

## What kind of answer I'm looking for

<fact? direction? a judgment criterion? — most open questions worth keeping want a framework, not
a fact>

---

# Sub-questions

- <optional: concrete sub-questions this decomposes into>

---

Upstream:
- <what triggered this>

Downstream (expected):
- <what it would feed if resolved — Insight / Hypothesis / Principle, or a specific document section>

Status semantics (optional): open (exploring) / inactive (shelved) / resolved (formed a conclusion — usually means it should be promoted, see Lifecycle below).

Lifecycle: a Question's typical fate is → Insight (suddenly clear) → Hypothesis (formed a testable judgment) → Principle (validated over time) → Dropped. Being "answered" is not the only successful outcome — dropping it because it stopped mattering is also fine; say so rather than forcing an artificial resolution.

7. Report back (all types)

Give the note's path and a one-line summary of the core content (Before/After for Insight, the decision statement for Principle, the one-sentence claim for Hypothesis, the question itself for Question, the core idea for Thought Mode, the one-sentence "what you'd build/do" for Idea) — don't re-paste the whole note. If step 0's classification differed from what the user asked for, restate which type it actually landed as and why, in the same message. If §0.5 routed the note to a vault other than the main vault, say which one and why, in the same message.