respect

respect

Teach Claude once — it remembers forever.

A feedback system for Claude Code that turns corrections into executable guard scripts and tips into persistent learnings — across every session and every project.

Install in 30 seconds

claude plugin marketplace add NixoN2/respect
claude plugin install respect@respect

Restart Claude Code. Requires jq on your PATH.

What Makes This Different

🧠 AI-Powered Learning Extraction Write natural feedback like "great job on the github workflow, you updated the repo first then bumped the version" — Claude analyzes the conversation context and extracts a structured learning with lesson, context, and reasoning. Generic praise stays as appreciation. Only actionable feedback becomes a durable learning.

🛡️ Guard Scripts — Executable Mistake Prevention /oops doesn't just record what went wrong — it generates a guard script that programmatically detects the mistake before it happens. Date checks, grep patterns, file existence — real code that runs on every action via a PreToolUse hook.

⚠️ Guard [no-timeout]: timeout is not available on macOS - use gtimeout
   Lesson: Don't use timeout command on macOS
   Action blocked.

Guards work on Bash commands, file writes, and MCP tools — target Slack, Datadog, GitHub, or any custom MCP server with regex patterns.

🌍 Cross-Project Memory Claude decides if each learning is project-specific or universal. Universal lessons (like "don't use timeout on macOS" or "push repo before bumping version") are stored globally and loaded in every project. Teach once, apply everywhere.

📊 Full Analytics Suite Six slash commands for the complete feedback loop:

Command	What it does
/tip	Reward good work. Natural language — size keyword anywhere in the message
/oops	Correct a mistake. Generates guard scripts + persistent learnings
/respect-stats	Tier progress, tip/correction ratio, trends, top patterns
/respect-export	Export learnings as portable CLAUDE.md — carry training to any project
/respect-guards	List, test, fix, or delete guard scripts
/respect-setup	Configure tiers, thresholds, tip sizes, and emoji

🏆 Tier Progression Earn points to climb tiers. A visible record of the trust you've built.

Tier	Balance
lurker	0	👤
contributor	20	🌱
trusted	60	⚡
veteran	150	🔥
partner	300	🤝

🎉 First-Run Onboarding New users get a welcome message explaining the system on their first session. No setup required — just install and start using /tip and /oops.

Quick Start

Tip when Claude does something well:

/tip large great work on the flexible parsing
/tip medium you handled the edge cases perfectly
/tip custom 10 amazing architecture decision

Correct when Claude makes a mistake:

/oops medium you forgot to bump the version
/oops large you used timeout which doesn't exist on macOS
/oops custom 5 missed the cached plugin directory

The size keyword (small, medium, large, custom N) can appear anywhere in your message. Claude extracts it intelligently.

Check your progress:

/respect-stats
/respect-export
/respect-guards

How It Works

Four systems working together:

1. Wallet — wallet.json tracks balance, tier, and history. Scripts update atomically — Claude can't modify it directly.

2. Learning Extraction — Claude analyzes conversation context to extract structured learnings from your feedback. Only actionable insights survive. Stored per-project in auto-memory, with universal lessons also saved globally.

3. Guard Scripts — /oops generates executable bash scripts that run on every tool action. Each guard has a trigger file (keywords + tool patterns) and a script (the actual check). Guards cover Bash, Write, Edit, NotebookEdit, and any MCP tool.

4. Cross-Project Memory — Claude evaluates each learning's scope. Project-specific lessons stay local. Universal lessons are stored in global-feedback.md and loaded in every project at session start.

Guard Script Examples

"Don't use timeout on macOS" — fires when Bash input contains "timeout":

{"triggers": ["timeout"], "lesson": "Use gtimeout on macOS instead"}

"Don't deploy on Fridays" — checks the day before git push:

{"triggers": ["git push", "deploy"], "lesson": "Avoid Friday deploys"}

"Don't post to Slack outside work hours" — targets Slack MCP tools:

{
  "triggers": ["send_message"],
  "tools": ["mcp__.*slack.*"],
  "lesson": "Post during work hours only"
}

The tools field accepts regex patterns. Omit it to default to Bash/Write/Edit/NotebookEdit. Use mcp__.* to match all MCP tools, or target specific servers like mcp__.*datadog.*.

What Becomes a Learning?