coach

name	coach
description	Use when Claude makes repeated mistakes, users correct behavior, tool failures indicate missing knowledge, or managing learning via /coach commands.
license	(MIT AND CC-BY-SA-4.0). See LICENSE-MIT and LICENSE-CC-BY-SA-4.0
compatibility	Requires python3 (for signal detection, aggregation, and analysis scripts).
metadata	{"author":"Netresearch DTT GmbH","version":"2.6.0","repository":"https://github.com/netresearch/claude-coach-plugin"}
allowed-tools	Bash(python3:*) Bash(python:*) Bash(sqlite3:*) Read Write Glob Grep

Coach - Self-Improving Learning System

Coach enables Claude to learn from friction and improve over time. It detects learning opportunities (user corrections, repeated instructions, tool failures, tone escalation), extracts actionable improvement candidates, and proposes changes requiring explicit user approval.

Core Principle: No silent writes. All improvements require user approval via /coach approve.

Activation Triggers

Activate when: user corrections ("no", "stop", "don't"), repeated instructions, tool/command failures, tone escalation (ALL CAPS, "!!!"), skill supplements ("also remember..."), deprecated-tool warnings, explicit /coach commands, or session end.

Signal Categories (Priority Order)

1. COMMAND_FAILURE (Highest) - Non-zero exit, stderr patterns
2. USER_CORRECTION (High) - Explicit correction language
3. SKILL_SUPPLEMENT (High) - Additional guidance for a skill
4. VERSION_ISSUE (Medium-High) - Deprecated/outdated warnings
5. REPETITION (Medium) - Semantically similar instruction repeated
6. TONE_ESCALATION (Low) - Frustration indicators

Candidate Types

Type	Description	Example
rule	Stable constraint	"Never edit generated files"
checklist	Workflow step	"Run tests after code change"
snippet	Repeatable command	"Preflight check script"
skill	Skill update suggestion	"Add X guidance to Y skill"
antipattern	Things to never do	"Never assume tool exists"

Workflow Summary

1. Detection — hooks capture events → ~/.claude-coach/events.sqlite
2. Generation — aggregate signals into proposals (fingerprints dedupe)
3. Scope — project vs global per path/language
4. Review — /coach review approves/rejects/edits
5. Apply — approved rules → ~/.claude/CLAUDE.md (global) or repo AGENTS.md (project)
6. Retro — /coach retro maps manual work to skill gaps, opens PRs at source repos

File Locations

~/.claude-coach/
├── events.sqlite      # Raw friction events
├── candidates.json    # Pending proposals
└── ledger.sqlite      # Cross-repo fingerprints

~/.claude/CLAUDE.md     # Global rules destination
<repo>/AGENTS.md        # Project rules destination

~/.claude/ or <repo>/.claude/
├── checklists/        # Workflow checklists
└── snippets/          # Reusable commands

Scope

Coach: live in-session signal — events.sqlite capture, /coach commands. retro: session-end / cross-session analysis, skill-PR routing. /coach retro overlaps retro's job; merging them is an open decision.

Scripts

scripts/ (this dir) — self-contained for skill-dir installs. Plugin path: ${CLAUDE_PLUGIN_ROOT}/skills/coach/scripts/:

Script	Purpose
init_coach.py	Initialize coach system
detect_signals.py	Pattern detection for friction
aggregate.py	Turn signals into candidates
skill_analyzer.py	Analyze skills and scan for outdated tools
apply.py	Apply approved proposals

Proactive Commands

• /coach scan — Check for outdated CLI tools and dependencies
• /coach retro — Session retrospective: analyze manual work, create PRs to improve skills at source

For detailed architecture, schemas, and patterns, see references/ directory.