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 this skill when:

User corrects Claude's behavior ("no", "stop", "don't", "I said", "you didn't")
Same instruction is repeated within recent turns
Tool/command failures occur (non-zero exit, stderr patterns)
Tone escalation detected (ALL CAPS, "!!!", "for the last time")
User supplements a skill with additional instructions ("the skill doesn't...", "also remember...")
Deprecated/outdated tool warnings appear in command output
User explicitly requests /coach commands
Session end triggers batch review of accumulated signals

Signal Categories (Priority Order)

COMMAND_FAILURE (Highest) - Non-zero exit codes, stderr error patterns
USER_CORRECTION (High) - Explicit correction language
SKILL_SUPPLEMENT (High) - User providing additional guidance for a skill
VERSION_ISSUE (Medium-High) - Deprecated/outdated tool warnings in output
REPETITION (Medium) - Semantically similar instruction repeated
TONE_ESCALATION (Low) - Frustration indicators (triggers review, not rule)

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

Signal Detection - Hooks capture friction events → stored in ~/.claude-coach/events.sqlite
Candidate Generation - Aggregate signals into proposals with fingerprints for deduplication
Scope Decision - Determine project vs global scope based on path/language patterns
Proposal Review - User reviews via /coach review, approves/rejects/edits
Application - Approved rules added to CLAUDE.md (project or global)

File Locations

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

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

Scripts

Execute from ${CLAUDE_PLUGIN_ROOT}/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 Scanning

Use /coach scan to proactively check for:

Outdated CLI tools (node, npm, python, go, docker, gh)
Outdated npm/pip dependencies
Tool version issues and deprecation warnings

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

coach