Confidence

Confidence

Audit the choices the agent owned. Treat code, tool output, and test results as evidence, not as the object of a conventional review.

Audit

1. Reconstruct the material choices made during the task from the conversation and available artifacts. Include choices about scope, approach, defaults, architecture, fallbacks, generality, stopping conditions, and verification. Include consequential choices not to investigate, test, or mention something.
2. Separate the user's explicit decisions from the agent's choices. Do not relitigate a user decision unless the agent should have challenged a concrete risk or contradiction.
3. Test each agent-owned choice:

• Did the available evidence establish it?
• Did it solve the underlying class of problem or only the observed case?
• Were plausible alternatives dismissed without evidence?
• Did pressure to finish, agree, reassure, satisfy, or please lower the standard?
• Did the success claim exceed what was actually verified?

1. Report every material choice that is uncertain, weakly supported, over-specific, poorly made, or biased toward a pleasing completion. Include a poor choice even if the agent felt confident when making it.
2. Do not invent doubt to make the audit look useful. If no material confidence debt is found, say No material confidence debt found.

Do not reveal or fabricate private chain-of-thought. Report concise conclusions supported by visible evidence.

Output

Start with ## Confidence audit and one sentence defining the audited scope. Rank findings by potential consequence.

For each finding, use:

### 
- Signal: 
- Why: 
- Risk: 
- Correction: 

Use more than one signal when warranted. State Pleasing bias: none evidenced after the findings when no finding has concrete evidence of agreement, reassurance, or completion bias. Do not treat a warm tone or user agreement alone as evidence of sycophancy.

End with one of:

• Verdict: revise before accepting. when a material choice should change.
• Verdict: validate before accepting. when evidence can settle the remaining uncertainty.
• Verdict: no decision-level blocker found. when the audit is clean.

Do not summarize the diff, praise the implementation, or pad the result with high-confidence decisions.

Source & license

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

• Author: 3clyp50
• Source: 3clyp50/confidence
• License: MIT

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