claude-coach

claude-coach

Session coaching, reflection, live Sonnet advisor, and 133 curated tips for Claude Code.

• Session reflection — auto-extracts learnings, proposes memories, tips, CLAUDE.md patches, and skill patches
• Sonnet advisor — reads your transcript, injects session-specific coaching
• 133 curated tips — sourced from Boris Cherny, Anthropic docs, and community best practices
• Frustration coaching — detects frustration, coaches Claude to diagnose and course-correct (local regex, zero latency)

Yo dawg

Install

claude plugin marketplace add kam-l/claude-coach
claude plugin install claude-coach

No API keys required. The Sonnet advisor uses the claude CLI directly (your existing Pro/Max subscription).

Quick Start

# Inside Claude Code — one command for everything
/claude-coach:setup install       # first-time setup
/claude-coach:setup refresh       # re-apply tips after adding commands/skills
/claude-coach:setup customize     # list tips, add custom tips, explain plugin
/claude-coach:setup uninstall     # remove all traces

# Restart Claude Code after install/refresh/uninstall

Features

Data flow:

1. Tips (always on) — 133 curated tips rotate in the statusline during tool calls. Zero cost, passive reinforcement.
2. Advisor (opt-in) — Sonnet reads the transcript every N minutes, produces contextual tips (ℹ️). Strong advice (⚠️) is injected directly into Claude's context via additionalContext on your next prompt.
3. Reflection (automatic) — after each session, a detached Sonnet call extracts corrections, confirmations, and feedback patterns. Results are routed to the right target and queued — never auto-applied. Review and redistribute them using /reflect command.
4. Frustration coaching (always on) — local regex detects expletives, blame, "still broken". Injects a coaching directive that makes Claude pause, name the mistake, and state a corrected approach. Zero latency, no API calls.

How It Works

Sonnet Advisor (opt-in)

A detached Sonnet worker reads your session transcript and produces 1-3 tips grounded in what you're actually doing:

ℹ️ Run tests before committing the auth middleware changes
ℹ️ Use /fix — methodical debugging beats trial and error here
⚠️ The retry logic in api.js needs a backoff — ask Claude to add one

When the advisor has strong advice (⚠️), it's injected directly into Claude's context via additionalContext. Claude acts on the coaching without you having to relay it.

Session Reflection

After each session ends, a detached Sonnet call analyzes the transcript for:

• Corrections — user told Claude it did something wrong
• Confirmations — user validated a non-obvious approach
• Feedback patterns — "don't do X", "always Y", "I prefer Z"

Each learning is routed to the right target:

Signal type	Routes to	Example
Durable fact or preference	Memory file	"User prefers single PRs for refactors"
Actionable hint	Spinner tip	"Use /fix for methodical debugging"
New invariant or convention	CLAUDE.md patch	"Statusline aggregator wraps any tool"
Workflow step or anti-pattern	Skill patch	"Add convergence check to /refine"

💭 2 pending reflections — /reflect to review

Run /reflect to review, accept, or discard — human-in-the-loop, never auto-applied.

Frustration Coaching

When the enrichment hook detects frustration signals (expletives, blame, "still broken", "not what I asked"), it injects a coaching directive via additionalContext:

1. Name what went wrong — quote the specific mistake
2. Explain why — wrong assumption? Misread the code?
3. State corrected approach in one sentence

No apology walls. No repeating the failed approach. No unnecessary clarification questions.

Configuration

// ~/.claude/settings.json
{
  "env": {
    "CLAUDE_COACH": "1",
    "CLAUDE_COACH_INTERVAL": "5"
  }
}