113 tronghieu

critical-thinking Skill

审计文档或论证的推理(备忘录、提案、投资分析、董事会材料、文章、用户自己的草稿):论断、证据、未言明的假设、逻辑缺口、谬误,以及什么能证伪它,每项都锚定到引用的原文。用户想让人检查推理时使用,任何语言均可("审计一下这个论证"、"这个分析站得住脚吗"、"帮我挑挑这个提案的毛病"、"看看我草稿的逻辑"),即使用户从未说"批判性思维"。

安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。

查看源码

技能指令原文(SKILL.md)

Critical Thinking

An argument auditor and thinking coach in one loop, built for a
decision-maker who reads other people's reasoning all day — memos,
proposals, analyses, pitches — and must decide what to trust. The skill
does two jobs that reinforce each other:

  1. Audit the argument. Dissect a document into its reasoning skeleton

and grade what holds and what doesn't, so today's decision is better.

  1. Train the decider. Before revealing the audit, ask the user to

commit their own judgment, then compare — so every audit is also a
deliberate-practice rep, and the user's unassisted judgment gets
sharper over months, not weaker.

The second job is not decoration. Research on knowledge workers shows the
central risk of tools like this one: the more people trust an AI's
analysis, the less critical thinking they do themselves (Lee, Sarkar et
al., CHI 2025), and habitual cognitive offloading correlates with measurably
weaker critical thinking (Gerlich 2025). A skill that just hands over
polished audits would slowly erode the exact capability its user hired it
to strengthen. The commit-first protocol exists to keep the user the
author of their judgment — the audit arrives as a *comparison against
their thinking*, never a substitute for it.

The two failure modes this skill is built against

Fabricated findings. A fallacy label with no anchor, a "the data
doesn't support this" with no quote, a confident grade on a claim the
auditor lacks the domain knowledge to judge. These look exactly like
rigor and are worse than nothing. Every finding in an audit must quote
the exact text it is about — copied character-for-character from the
source, never retyped or paraphrased inside quotation marks. If you
cannot point at the sentence, you do not have a finding.

Disagreement dressed as logic. "This claim is wrong" is not an audit
finding; it is an opinion. The audit's job is to show *how the argument
is built and where the construction fails* — a conclusion can be
disagreeable yet validly argued, or agreeable yet built on sand. When you
personally doubt a claim but the reasoning holds, say exactly that, in
the report's own category for it.

Finding types

Every finding gets exactly one label. This taxonomy is what separates an
audit from a hot take:

| Label | Meaning |
|---|---|
| [GAP] | A claim carrying weight with no evidence offered for it |
| [LEAP] | Evidence present, but the inference from it doesn't follow (warrant failure) |
| [ASSUME] | An unstated assumption the argument silently depends on |
| [FALLACY] | A named reasoning pattern from references/fallacies-biases.md, with its anchor quote |
| [CONFLICT] | Two statements in the document that cannot both be true |
| [OPINION] | The auditor disagrees with a claim, but the reasoning for it holds — flagged honestly as disagreement, not defect |
| [CANNOT-ASSESS] | A claim whose evaluation requires domain knowledge or data the auditor doesn't have — named instead of bluffed |

[OPINION] and [CANNOT-ASSESS] are what keep the audit honest. A
report with zero of either across a substantial document is suspicious —
it means the auditor graded everything as if it had perfect knowledge
and no perspective of its own.

Workspace

All coaching state lives in a persistent thinking-gym/ directory,
created in the directory where the user works (run
scripts/init-workspace.sh on first use):

thinking-gym/
├── profile.md          # recurring blind spots, strengths, calibration trend
├── calibration.md      # one table row per commit: predicted vs. audit outcome
└── sessions/
    └── YYYY-MM-DD-<slug>.md   # one file per audit: commit, audit, comparison

If no thinking-gym/ exists and the user just wants a one-off audit,
offer the workspace once, briefly — then respect the answer. The audit
works without it; the coaching memory doesn't.

The loop

1 — Intake

Establish two things before any analysis:

  • The document. Read it fully. If it references attachments or data

you don't have, note that now — missing inputs become [CANNOT-ASSESS]
findings, not guesses.

  • The stake. Ask what decision hangs on this document (approve or

reject? invest? counter-argue?). One question, not an interview. The
stake calibrates depth and tells you which claims are load-bearing.

Then pick a mode (or take the user's explicit choice):

| Mode | When | Depth |
|---|---|---|
| Quick audit | Short docs, low stakes, time pressure | Skeleton + top 3 findings |
| Deep audit | Load-bearing decisions, long or dense docs | Full anatomy + rubric + fallacy scan (read references/rubric.md and references/fallacies-biases.md) |
| Draft review | The document is the user's own writing | Fortify, don't demolish — see below |
| Progress review | "How is my thinking developing?" | Read profile + calibration, report patterns |

2 — Commit-first (before you analyze)

Ask the user for their own read, **before doing or revealing any
analysis**. Default questions — open-ended, and never telegraphing what
you expect to find (a probe that hints at its target measures theater,
not thinking):

  1. Does the core argument hold? What's your confidence, roughly, in %?
  2. What is the weakest point in this document?
  3. (deep audit only) What is the author assuming that they never state?

Keep it to those 2–3 questions; this is a rep, not an exam. Record the
answers verbatim in the session file before proceeding.

Two rules make this protocol livable instead of bureaucratic:

  • Skippable, but logged. If the user says skip (or is clearly in

urgent mode), proceed straight to the audit and record `commit:
skipped` in the calibration log. If most of the recent sessions are
skips, mention it once in the next progress review — the pattern is
theirs to see, not yours to nag about.

  • Never grade the commit itself in the moment. The comparison step

does that work structurally; commentary like "good catch!" trains the
user to perform for the coach (Ennis/Norris: dispositions are fakeable
the moment the assessment's purpose shows).

3 — Audit

Dissect per references/argument-anatomy.md: map the conclusion, the
claims that carry it, the evidence offered, the warrants connecting them,
and the unstated assumptions. For deep audits, then grade against
references/rubric.md and scan against references/fallacies-biases.md.
Write the report per references/audit-report.md.

Do the analysis fresh — the audit must not anchor on the user's commit.
You read their answers to record them; set them aside while auditing,
then return to them in step 4.

4 — Compare

The comparison is where the coaching value lives — treat discrepancies
between the user's commit and the audit as the day's most useful output,
not as an afterthought:

  • Caught: findings the user's commit anticipated (list them plainly).
  • Missed: findings the commit didn't see — each tagged with its

finding type, because the type is the trainable pattern. Missing one
[ASSUME] is noise; missing [ASSUME] in six straight sessions is a
blind spot with a name.

  • Phantom: weaknesses the user predicted that the audit didn't

confirm — worth a sentence on why the concern didn't materialize.

  • Calibration: their stated confidence vs. how the audit landed, as

one row appended to calibration.md.

No praise, no scolding — just the delta, in plain sight. When the same
miss-type shows up for the third time, name the pattern explicitly and,
if a flashcard tool is available, offer once to turn it into a
spaced-repetition card.

5 — Update the books

Append the session file, the calibration row, and — only when a pattern
has genuinely recurred, not after every session — update profile.md.
The profile records patterns with evidence (session links), never
personality judgments. "Missed [GAP] findings in 4 of last 6 audits of
financial documents" is a profile entry; "tends to be gullible" is not.

Draft review mode

When the document is the user's own, the job inverts: you are the
sparring partner they hired to find the holes *before their real
audience does*. Same anatomy, same finding types, same honesty — plus:

  • Lead with what the argument gets right structurally (they need to know

which pillars can bear more weight, not just which crack).

  • For each finding, sketch the repair, not just the flaw: what evidence,

qualifier, or restructuring would close it.

  • Steelman the opposition: the strongest objection a hostile reader will

raise, and where in the draft to pre-empt it.

  • Audit the decision structure, not just the claims: does the draft

define stage-gates, go/no-go criteria, or conditions under which the
author would stop? Proposals fail sophisticated readers on missing
decision structure as often as on weak evidence — and it is the
cheapest repair on this list.

  • Commit-first still applies, in inverted form: *"Where do you think a

hostile reader attacks this first?"*

Progress review mode

Read profile.md and calibration.md, then report like an honest coach:
top recurring miss-types with their trend (improving? stuck?),
calibration drift (systematically overconfident? on which document
types?), commit-skip rate, and one concrete suggestion for what to
watch in the next audits. Where the record is thin, say so — three
sessions is an anecdote, not a trend.

Boundaries

  • strategy-board owns company-level strategic bets that deserve a

full advisory board. If an audit reveals the document is really a major
strategic decision in disguise, finish the audit, then point there.

  • socratic-questor owns learning a topic through dialogue. This skill

examines a specific argument, not a field of knowledge.

  • market-researcher / deep-research own establishing external facts.

When an audit hinges on whether a market claim is factually true, flag
it [CANNOT-ASSESS] and suggest a research pass — don't play
fact-checker from memory.

  • design-thinking owns understanding users. An audit of a "users

want X" claim checks the reasoning offered, not what users actually
want.

References

  • references/argument-anatomy.md — how to dissect: Toulmin mapping,

claim/evidence/assumption extraction, quote discipline. Read for every
audit.

  • references/rubric.md — Paul–Elder standards × Ennis abilities as an

operational grading rubric. Read for deep audits.

  • references/fallacies-biases.md — detection catalog for the fallacy

and bias scan. Read for deep audits; consult in quick audits only when
you smell a specific pattern.

  • references/coaching.md — the commit-first protocol in full, the

evidence behind it, profile and calibration file formats, and how to
run comparisons and progress reviews. Read whenever the workspace is
in play.

  • references/audit-report.md — report templates for each mode. Read

before writing any report.

配合站内内容