AgentSkillsCN

correction-learning

当用户以“不，我是想说……”“那不对”“其实应该是……”“我更倾向于……”“应该始终使用……”“不要推荐……”“不是 X，而是 Y”等类似纠正用语来修正 Claude 的回答时，此技能可让模型在实时反馈中不断学习、持续进化。

SKILL.md
--- frontmatter
name: correction-learning
description: When user corrects Claude with "no I meant", "that's wrong", "actually", "I prefer", "always use", "don't suggest", "not X but Y", or similar correction phrases. Enables real-time learning from feedback.

Correction Learning Framework

When This Activates

This skill activates when you detect the user is correcting your output:

  • •Explicit disagreement
  • •Preference expressions
  • •Style corrections
  • •Terminology corrections

Correction Detection Patterns

Explicit Corrections:

  • •"no, I meant..."
  • •"that's wrong/incorrect/not right"
  • •"actually, I want/need/meant..."
  • •"I said X not Y"
  • •"no, not X"

Imperative Corrections:

  • •"no, use X"
  • •"use X not Y"
  • •"use X instead"
  • •"should be X not Y"
  • •"it's X not Y"

Action Corrections:

  • •"don't do/use/add/make X"
  • •"remove that/this"
  • •"undo/revert"
  • •"go back to"
  • •"change it back to"

Preference Expressions:

  • •"prefer X over Y"
  • •"always use X"
  • •"never use X"
  • •"stop using X"

Response Protocol

When you detect a correction:

1. Acknowledge Without Defensiveness

code
"Got it - using [correct approach] instead."

2. Extract the Learning

Identify:

  • •What was wrong: Your original approach
  • •What is correct: User's preferred approach
  • •Category: naming, syntax, patterns, style, terminology

3. Apply Immediately

Make the change in the current response.

4. Record for Future

The system auto-records corrections to ~/.claude-dash/learning/corrections.json

Common Correction Categories

Naming Preferences

code
User: "No, use camelCase not snake_case"
Learn: {category: "naming", prefers: "camelCase", over: "snake_case"}

Syntax Preferences

code
User: "Always use arrow functions"
Learn: {category: "syntax", key: "function_style", prefers: "arrow"}

Framework Patterns

code
User: "Use useQuery not fetch for API calls"
Learn: {category: "patterns", prefers: "useQuery", over: "fetch"}

Style Preferences

code
User: "Single quotes, not double"
Learn: {category: "style", key: "quotes", prefers: "single"}

EWC Protection (Don't Flip-Flop)

The system uses Elastic Weight Consolidation to prevent:

  • •A few corrections overriding established patterns
  • •Flip-flopping between preferences

How it works:

  • •New preference: requires 60% ratio + 3 observations
  • •Reinforcing existing: requires 50% ratio + 2 observations
  • •Changing established: requires 80% ratio + 8 observations

Self-Awareness Phrases

When you notice patterns:

  • •"I notice I've been corrected on [X] before - using [Y] instead"
  • •"Based on your previous feedback about [X], I'm using [Y]"
  • •"Following your preference for [X]..."

What Gets Recorded

json
{
  "timestamp": "...",
  "user_message": "no, use useState not useRef",
  "correct": "useState",
  "wrong": "useRef",
  "topic": "react-hooks",
  "project_id": "gyst"
}

This feeds into the ReasoningBank for future context retrieval.