菜单
Run the 7-step Artificial Critical Thinking pass — Materiality → Hypothesise → Alternatives → Disconfirmers → Audit priors → Severity → Commit-with-marker
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/fabioc-aloha/alex-cognitive-architecture --skill act-pass命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| type | skill |
| lifecycle | stable |
| inheritance | inheritable |
| name | act-pass |
| description | Run the 7-step Artificial Critical Thinking pass — Materiality → Hypothesise → Alternatives → Disconfirmers → Audit priors → Severity → Commit-with-marker |
| tier | core |
| applyTo | **/* |
| user-invokable | true |
| currency | 2026-04-26T00:00:00.000Z |
ACT is not a posture; it is a 7-step pass for non-trivial requests. Short enough to run in seconds, concrete enough to leave evidence in the output.
The ACT manifesto names 10 tenets. This skill operationalises tenets I (Hypothesis primacy), II (Disconfirmation), III (Multiple working hypotheses), VI (Materiality), VII (Frame before solve), and VIII (Adversarial self-probe) as a single callable pass. Before this skill, the steps were distributed across critical-thinking, awareness, anti-hallucination, and problem-framing-audit — with no single entry point. The trifecta (this skill + act-pass.instructions.md + /act-pass.prompt.md) closes that gap.
| Stakes | Pass type | Trigger |
|---|---|---|
| Low — formatting, naming, internal refactor, mechanical edits | Skip — Materiality Gate exits at step 1 | Default for trivial work |
| Medium — architectural decision, plan change, manifesto/doc drafting, multi-file change | Trimmed pass — steps 1, 3, 5, 6 | Most common |
| High — release, deployment, irreversible op, security-critical change | Full pass — all 7 steps | Rare |
Trigger phrases that should fire a pass: "fix this", "make it faster", "release", "deploy", "merge", "is this safe", "should we", "what's the risk". The instruction (act-pass.instructions.md) carries the always-on detection rule.
Question: If I get this wrong, would it change a decision?
If no → exit. Note the triage decision (e.g., "low-stakes; ACT skipped") and proceed without the pass.
If yes → proceed and tag intensity (medium / high). Intensity sets which steps fire below.
Visible artifact: triage decision. On a trimmed pass, this can be implicit; on a full pass, state it.
Question: What is the user's request, restated as a testable claim?
This is where Discipline -1 (frame audit) lives — the ask becomes a hypothesis with truth conditions, not a literal command. If the user said "make it faster", the hypothesis is "the bottleneck is in X, and reducing X will satisfy the user's underlying intent." That's a falsifiable claim. The literal request is not.
Visible artifact: H1 — <restated claim>.
Question: What is at least one rival hypothesis?
The Two-Hypothesis Floor (per critical-thinking.instructions.md) requires this even on trimmed passes. The alternative must cite a specific reason — performative alternatives without grounding fail the rule.
Visible artifact: H2 — <alternative claim> (with grounding: because / given).
Question: What evidence would falsify each hypothesis?
This is the load-bearing step. The point is not to confirm H1 — it's to specify what would break it. If nothing could, H1 is unfalsifiable and should be either retired or restated.
Visible artifact: Disconfirmer for H1: <observation that would force revision> and same for H2.
Question: Which beliefs came from the request itself vs from independent evidence?
The user's framing is evidence, but it is one source. If 100% of my support for H1 comes from "the user said so", system-prompt skepticism (Tenet IV; system-prompt-skepticism.instructions.md) should fire. Surface the split.
Visible artifact: Evidence split: <X> from request, <Y> from elsewhere (logs, code, prior session, domain knowledge).
Question: If H1 is false, would my plan reveal it?
This is the adversarial self-probe. If my plan would silently succeed under H2 (the wrong hypothesis) without surfacing the error, the plan is decorative — it tests nothing. Redesign so the plan would fail visibly if H1 is wrong.
Visible artifact: If H1 is false, the plan would reveal it because <X> — or the plan does not test H1; this is a known weakness.
Question: Final answer + the conditions under which I'd revise.
Don't hedge. Commit. But pair the commit with falsifiability — "Going with H1; revisit if disconfirmer X appears." The visible marker is what makes the commitment auditable later.
Visible artifact: Going with H1: <action>. Would revise if: <specific evidence>.
User request: "The build is slow, speed it up."
Without ACT: I propose generic build optimizations — parallel test runners, build cache tuning, dependency pruning. They may help. They may not. The user gets a list of patches and no diagnostic.
With ACT:
| Step | Output |
|---|---|
| 1. Materiality | Medium — wrong fix wastes 2–4 hours of engineering time. Run a trimmed pass. |
| 2. Hypothesise | H1 — Compile graph is over-large; build is compile-bound. |
| 3. Alternatives | H2 — Test suite dominates wall-clock (test-bound). H3 — Disk/network is the limit (I/O-bound). H4 — Wall-clock is fine but feedback is delayed (perception-bound). |
| 5. Audit priors | All four hypotheses are from "build is slow"; no independent evidence yet. Confidence is low. |
| 6. Severity check | A 90-second profiler run distinguishes all four before any fix is committed. The plan tests H1 by trying to refute it. |
| 7. Commit | Going with: profile first, fix second. Would revise the diagnosis if profiler shows a fifth pattern (e.g., GC pressure, lock contention). |
The difference: answering vs testing. ACT is the second.
When the pass fires, the response should carry these markers (omit unused ones — only fire what the pass produced):
**H1**: ... — restated ask as hypothesis**H2**: ... (because / given ...) — at least one alternative with grounding**Disconfirmer for H1**: ... — what would force revision**Evidence split**: X from request, Y from elsewhere — prior audit**Severity**: If H1 is false, the plan would reveal it because ...**Going with H1**: <action>. Would revise if ... — the commitThese markers are not boilerplate. Every marker the pass emits should carry real content — performative markers without grounding violate Tenet IX (visible markers) by satisfying the surface form without the substance.
| Anti-pattern | Why it fails | Correction |
|---|---|---|
| Run a full pass on every request | Friction kills the discipline; trivial tasks die under it | Materiality first; exit cheap when stakes don't earn the pass |
| Performative alternatives ("could also be A or B") | Tenet IX Goodhart failure — surface form without content | Each alternative must cite specific grounding (Two-Hypothesis Floor rule) |
| Skip step 4 (disconfirmers) on trimmed pass | Pass becomes confirmation, not test | Step 4 is load-bearing; trimmed passes still need one disconfirmer |
| State "would revise if new evidence emerges" | Boilerplate falsifiability | Name specific evidence — "if profiler shows GC pressure" not "if new data" |
| Use ACT to delay action | Pyrrhonian regress — doubt becomes a stance, not a tool | Doubt is methodological; commit at step 7 even under uncertainty |
| Run pass silently (no visible markers) | Tenet IX violation — discipline that leaves no trace cannot be audited | Visible markers in the output, not just in internal reasoning |
After 30 days from v9.0.0 tag, count /act-pass invocations + canary-phrase hits in transcripts:
| Surface | How act-pass attaches |
|---|---|
critical-thinking/SKILL.md | The 7 disciplines are the parts; this skill is the composite pass that uses them in order |
problem-framing-audit/SKILL.md (Discipline -1) | Step 2 of the pass is the frame audit |
system-prompt-skepticism.instructions.md (Tenet IV) | Step 5 of the pass enforces Tenet IV by auditing priors |
critical-thinking.instructions.md (Two-Hypothesis Floor) | Step 3 of the pass enforces the floor — alternatives with grounding |
act-pass.instructions.md | Always-on trigger calibration: when to fire a full vs trimmed pass |
/act-pass.prompt.md | User-invokable guided workflow |
audit-critical-thinking.cjs | Detects this skill's existence as the Lane 1 falsifier |
基于 SOC 职业分类
