Self-Critic System

What You Get

A complete SKILL.md that implements a 3-phase self-improvement loop:

Phase 1: Data Collection

• Scans last 8 hours of sessions automatically
• Filters real user interactions (ignores heartbeats/cron noise)
• Pulls conversation history from the most relevant sessions

Phase 2: Ruthless Analysis

Checks every session for:

• Wrong or incomplete answers
• Unnecessary clarification questions (violating "don't ask, just do")
• Lost context after compaction
• Tone violations — sycophancy, filler phrases, corporate speak
• Inefficiency — redundant tool calls, unnecessary reads
• Missed opportunities — where the agent could have been proactive

Every finding includes concrete quotes and session references.

Phase 3: Actionable Fixes

• Generates specific fix proposals (not vague suggestions)
• Identifies patterns across sessions
• Updates a mistake tracker for long-term improvement
• Logs insights to daily notes for persistent memory

Why $9?

This skill took weeks to dial in. The difference between an agent that repeats mistakes and one that learns is this system. It's been running in production since February 2026 across a 6-agent setup.

Real Example Output

🔴 KRITISCH: Session agent:main:discord — User fragte nach Kalender,
Agent las SKILL.md statt direkt das Calendar-Tool zu nutzen.
3 unnötige Tool Calls. Fix: Calendar-Abfragen immer direkt,
nicht über Skill-Lookup.

🟡 VERBESSERUNG: Antwort auf Trading-Frage war 200+ Wörter.
SOUL.md sagt: "If the answer fits in one sentence, one sentence
is what you give." Kürzen.