user-corrections

name	user-corrections
description	Captures and learns from direct user corrections to Claude Code's outputs. Complements the code-reviewer-driven learning system with user-driven feedback for continuous improvement.

User Correction Learning

This skill captures and learns from direct user corrections to agent outputs, enabling continuous improvement through user-driven feedback.

Overview

User corrections are a high-value learning signal. When a user corrects Claude's output, it indicates a gap between expected and actual behavior that should inform future work.

User corrects Claude → Agent detects correction → Logs to user-corrections.jsonl
                                                          ↓
                                    /evolve analyzes alongside review-findings
                                                          ↓
                                    High-confidence corrections → runbooks
                                                          ↓
                                    /corrections command for periodic review

Data Flow

┌─────────────────────────────────────────────────────────────┐
│                     DATA SOURCES                            │
├─────────────────────────────────────────────────────────────┤
│  .claude/review-findings.jsonl    (code-reviewer output)    │
│  .claude/session-learnings.jsonl  (any agent learnings)     │
│  .claude/user-corrections.jsonl   (user correction signals) │ ← NEW
│  .claude/incidents.jsonl          (production incidents)    │
└─────────────────────────────────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│                   PATTERN ANALYSIS                          │
├─────────────────────────────────────────────────────────────┤
│  Pattern extraction    → Group similar corrections          │
│  Source weighting      → Weight by correction source        │
│  Frequency counting    → Track occurrences over time        │
│  Confidence scoring    → Combine all signals                │
└─────────────────────────────────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│                   KNOWLEDGE OUTPUT                          │
├─────────────────────────────────────────────────────────────┤
│  .claude/patterns/patterns.json      (pattern registry)     │
│  .claude/patterns/trends.json        (trend analysis)       │
│  .claude/skills/runbooks/*/RUNBOOK.md (auto-updated)        │
└─────────────────────────────────────────────────────────────┘

Correction Types

Type	Definition	Example
code_quality	User fixes bugs, style, or logic	User fixes off-by-one error Claude introduced
code_completeness	User adds code Claude missed	User adds error handling Claude skipped
approach_rejection	User rejects overall approach	"Let's try a different approach"
expectation_mismatch	Claude did something unexpected	"I didn't ask for that"
communication_gap	Claude misunderstood request	User rephrases same request
preference_conflict	User prefers different pattern	"I prefer X over Y"

Commands

Command	Description
/corrections	Show recent corrections summary
/corrections --agent <name>	Filter by agent
/corrections --type <type>	Filter by correction type
/corrections --analyze	Run pattern extraction
/evolve	Analyze corrections alongside findings

Source Weighting

Corrections are weighted differently based on source quality:

Source	Weight	Rationale
Explicit user correction	1.0	Direct feedback, highest confidence
Implicit user correction	0.8	User action indicates issue
Blocking review finding	0.7	Code reviewer found issue
Warning review finding	0.5	Less severe but notable
Session learning	0.4	Agent self-observation
Nit review finding	0.3	Minor conventions

These weights apply to the source_quality_score component of confidence calculation.

Integration with Agents

During Session Work

Agents detect corrections in real-time:

## Correction Detection

When a user corrects your output:
1. Detect explicit signals ("wrong", "no", "actually...")
2. Detect implicit signals (user edits your code, re-requests differently)
3. Log correction to `.claude/user-corrections.jsonl`
4. Continue with corrected approach

During Context Discovery

Agents check for patterns from corrections:

## Context Discovery

...
N. Read `.claude/patterns/patterns.json` for known patterns
N+1. Note patterns with high correction-based confidence
N+2. Apply lessons from user corrections to current work

Session Learning Integration

User corrections feed into the same learning engine as review findings:

1. Raw correction → .claude/user-corrections.jsonl
2. Pattern extraction → /evolve groups similar corrections
3. Confidence scoring → Weighted by source quality
4. Runbook promotion → High-confidence patterns added to runbooks

Files

File	Purpose
SKILL.md	This overview
schemas.md	JSON schema for user-corrections.jsonl
detection.md	How agents detect and log corrections

Related Skills

• learning-engine — Pattern analysis and runbook promotion
• runbooks — Where learned patterns are stored