Skip to main content
The confidence-check skill prevents wasted effort from implementing with wrong assumptions. It runs a scored self-assessment across three dimensions and gates implementation on the result.

​
When this skill fires

The skill description reads: “Use before implementing a feature or making significant changes to verify you have enough context and understanding to proceed — prevents wasted effort from proceeding with wrong assumptions.” Specific triggers:
  • Before starting implementation of a feature
  • When requirements feel ambiguous
  • After reading existing code you’re about to modify
  • When you’re unsure about a technical approach

​
What it does

The skill scores three dimensions — Understanding, Context, and Approach — on a 0–10 scale. The total score determines whether to proceed, fill gaps first, or stop and discuss. An anti-gaming protocol prevents inflated scores: any dimension scored 8 or higher requires specific named evidence.

​
How it works

​
The three dimensions

Understanding (0–10)
  • Do you understand what this is supposed to do?
  • Do you understand why it needs to exist?
  • Do you understand who uses it and how?
Context (0–10)
  • Do you know which files you need to touch?
  • Do you understand the existing code patterns you’re building on?
  • Do you know the data flow end-to-end?
Approach (0–10)
  • Do you know which approach you’ll take and why?
  • Have you considered at least one alternative?
  • Do you know how to test this?

​
Anti-gaming protocol

For any dimension scored 8 or higher, you must write one specific piece of evidence naming a concrete artifact — a file, a line range, a function name, a specific behavior you observed. General statements do not count. If you cannot name specific evidence for a dimension scored 8+, the maximum score for that dimension is 6.
Valid evidence examples:
  • “Context: 9/10 — I have read auth.js:45-120 and understand the JWT validation flow, including how the middleware chain passes the decoded token to route handlers.”
  • “Understanding: 8/10 — The requirement says ‘users can reset their password via email.’ I know it means: generate a one-time token, email a reset link, validate the token on return, and expire it after 1 hour per the spec in JIRA-442.”
Invalid evidence (too vague):
  • “Context: 9/10 — I understand the codebase well.”
  • “Understanding: 8/10 — The requirement is clear.”

​
Scoring thresholds

​
External validation at 20–26

When total score is in the 20–26 range, state your understanding out loud in one paragraph — what you are about to build, why, and how. Be specific. If the user does not correct it within their next response, proceed. If they do correct it, re-run the check from the beginning with the updated understanding.

​
Context change handling

If your confidence drops below 20 mid-implementation: Stop immediately. Do not continue writing code. Announce: “Confidence has dropped to [X]/30 due to [specific reason]. Pausing implementation to [specific gap-filling action].” Then execute the gap-filling action. Do not resume until confidence returns to 20+ (with external validation) or 27+ for autonomous continuation.

​
What to do with low scores

Low on Understanding: Restate the requirement in your own words and ask the user to confirm. Do not assume you understood. Low on Context: Read the relevant files before proceeding. Do not write code that touches files you have not read. Use the deep-research skill for external unknowns. Low on Approach: Load the architecture-design skill. Explore 2–3 options before committing to one.

​
Example scenario

You’re about to implement a password reset endpoint. Understanding: 8/10 — Evidence: “I know the endpoint needs to: accept an email address, generate a time-limited reset token, send a reset link, validate the token on return, and allow the user to set a new password. Gap: I don’t know whether the token should be single-use.” Context: 6/10 — Evidence: “I have read user.model.js and understand the User schema. I have not yet read the email service. I can see it referenced in order.controller.js but don’t know its interface.” Approach: 7/10 — Evidence: “I will use crypto.randomBytes(32) for token generation, store the SHA-256 hash in a password_reset_tokens table with an expires_at column. Gap: email dispatch step is uncertain.” Total: 21/30 — Fill gaps before starting. Gaps:
  1. Read the email service implementation
  2. Confirm with user: single-use tokens?
State understanding and wait for confirmation before proceeding.

Deep research

Fills the Context gap when external unknowns (third-party APIs, library behavior) are the blocker.

Architecture design

Fills the Approach gap when the technical path is unclear.

Brainstorming

Required when total score falls below 20 — indicates the feature needs design before implementation.