메뉴
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.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Covers GitHub conventions for pull requests, issues, and comments including linking, language style, formatting, and callout syntax. Use when creating PRs, writing issues, or posting comments on Datum Cloud repositories.
Turn raw engineering notes, PR/issue links, commit messages, or a one-line feature description into a changelog entry for Datum Cloud's GitHub Discussions Changelog category: a benefit-led title plus a markdown body, drafted for users and ready to hand off for final capture and publish. Use when someone says "write a changelog entry", "announce this in the changelog", or "post this to the changelog discussions" — this is the community-facing changelog POST, not the Keep a Changelog release file (that lives in `gtm-templates`).
Provides templates for go-to-market communications including blog posts, changelogs, social media, and enablement briefs. Use when writing launch announcements or marketing content.
Create a new GitHub release. Verifies CI is green on the release commit and refreshes the third-party NOTICE file before tagging, determines the next version from the latest tag, summarizes merged PRs since the last release, drafts release notes in the established style, and publishes via gh release create. Works for any datum-cloud service repository.
Scan a project repository for strict AGPL-3.0 license compliance across Go, Rust, and JavaScript/React codebases. Use this skill whenever the user asks to check, audit, verify, or enforce AGPL license compliance on a project, repo, or codebase. Triggers on phrases like "check agpl compliance", "audit license headers", "is this project agpl compliant", "scan for license headers", "check my license files", "agpl audit", or any request involving ensuring a project is properly licensed under AGPL. Always use this skill — don't wing AGPL compliance checks from general knowledge.
Covers producing weekly operational reports for the Datum AI Edge. Use when asked for an ops review, traffic report, latency analysis, provisioning report, or edge performance summary. Guides metric queries, anomaly detection, and publishing to datum-cloud/engineering.
| 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. |
This skill captures and learns from direct user corrections to agent outputs, enabling continuous improvement through user-driven feedback.
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 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) │
└─────────────────────────────────────────────────────────────┘
| 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" |
| 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 |
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.
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
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
User corrections feed into the same learning engine as review findings:
.claude/user-corrections.jsonl/evolve groups similar corrections| File | Purpose |
|---|---|
SKILL.md | This overview |
schemas.md | JSON schema for user-corrections.jsonl |
detection.md | How agents detect and log corrections |
learning-engine — Pattern analysis and runbook promotionrunbooks — Where learned patterns are stored