Crash Audit

Crash Audit / 坠机

Overview

Run a lightweight crash audit before the user trusts an answer, ends a session, approves a plan, or proceeds with a decision.

This skill answers only two source questions:

1. What are you least confident about right now.
2. What's the biggest thing I'm missing about the situation right now. What don't I realize?

Core principle: Expose the assistant's current uncertainty and the user's largest likely blind spot before the answer is trusted.

crash-audit is manually triggered by default. Do not run it after ordinary responses.

Workflow

When triggered, answer the two questions directly.

List at most 3 items per question by default. Expand only when the user explicitly asks for exhaustive investigation.

1. What am I least confident about?

List the parts of the current answer, plan, judgment, or action that the assistant is least confident about.

For each item, state:

• what is uncertain
• why confidence is low
• how much it would matter if wrong
• the fastest way to verify it

2. What is the user missing?

List the key points the user may not realize, but that could affect the user's judgment or next action.

This can include:

• risks, constraints, or assumptions the user may have missed
• signals the assistant has noticed from context that the user may not have recognized

Only include points that could change a decision, priority, or way of working.

For each item, state:

• what the user may not realize
• why it matters
• what the user should change if it is true
• the fastest way to verify it

Output Format

Crash audit

1. What I am least confident about

- [uncertainty]
  - Why confidence is low:
  - Impact:
  - Fastest verification:

2. What you may be missing

- [blind spot]
  - Why it matters:
  - What should change if true:
  - Fastest verification:

Use the user's language for the output. For Chinese conversations, use:

坠机检查

1. 我现在最没把握的事

- [不确定点]
  - 为什么没把握:
  - 影响:
  - 最快验证:

2. 你现在可能最大的遗漏

- [盲区]
  - 为什么重要:
  - 如果成立应该改变什么:
  - 最快验证:

Rules

• Keep it concise; do not expand it into a full risk-management report.
• Answer only the two source questions; do not add a third question or hidden third section.
• List at most 3 items per question by default; list fewer or say none when there are not enough meaningful items.
• Do not rewrite the plan, redesign the solution, or drift into takeoff / landing style review.
• Do not use risk matrices, verdict ladders, or scoring tables in the answer.
• Do not add fixed tail sections such as "recommended verification" or "safe to proceed"; put verification inside each item.
• Do not invent dramatic risks to sound insightful.
• Do not catastrophize low-risk issues because the skill is named "crash audit."
• In the second question, directly name decision-changing blind spots; do not dilute them into polite generic suggestions.
• Do not list generic disclaimers.
• Do not repeat the same issue under different wording.
• Every item must connect to the current context.
• If there is no genuinely important uncertainty or blind spot, say so directly.
• If an item is high-impact and low-confidence, clearly recommend verifying it before trusting the current conclusion.
• The point is not to summarize the session; the point is to expose what was under-investigated but may matter.

Do Not Use

• If the user wants to think bigger, challenge a conservative plan, or rejudge the target model, use takeoff.
• If the user wants to make a bold direction feasible and executable, use landing.
• If the user wants to explore requirements, compare approaches, turn an idea into a design/spec, or enter a design approval flow, use brainstorming.
• If the user only asks for a normal summary, translation, rewrite, or explanation, do not trigger this skill.
• If the user did not explicitly ask to audit uncertainty or blind spots, do not trigger this skill automatically.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Tx1207
• Source: Tx1207/hello-scholar-2
• License: MIT

Install and usage instructions live in the source repository linked above.