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.
claude plugin marketplace add NixoN2/respect
claude plugin install respect@respectRestart Claude Code. Requires jq on your PATH.
🧠 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.
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
Four systems working together:
-
Wallet —
wallet.jsontracks balance, tier, and history. Scripts update atomically — Claude can't modify it directly. -
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.
-
Guard Scripts —
/oopsgenerates 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. -
Cross-Project Memory — Claude evaluates each learning's scope. Project-specific lessons stay local. Universal lessons are stored in
global-feedback.mdand loaded in every project at session start.
"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.*.
Structured Learnings (stored with Lesson/Context/Why):
- "You updated the repo first, then bumped version — perfect workflow"
- "You broke the tests by not running them before committing"
Simple Tips (stored as appreciation only):
- "Amazing work! You're the best!"
- "Thanks for the help, large tip"
| Command | Effect |
|---|---|
/tip small |
+1 pt |
/tip medium |
+3 pts |
/tip large |
+5 pts |
/tip custom N |
+N pts (max 10 by default) |
/oops small |
-1 pt |
/oops medium |
-3 pts |
/oops large |
-5 pts |
/oops custom N |
-N pts |
Add to ~/.claude/statusline.sh:
WALLET="$HOME/.claude/respect/wallet.json"
R_CONFIG="$HOME/.claude/respect/config.json"
if [ -f "$WALLET" ] && [ -f "$R_CONFIG" ]; then
R_BALANCE=$(jq -r '.balance // 0' "$WALLET")
R_NAME=$(jq -r '.tier // ""' "$WALLET")
R_EMOJI=$(jq -r --arg name "$R_NAME" '.tiers[] | select(.name == $name) | .emoji // ""' "$R_CONFIG")
R_FORMAT=$(jq -r '.statusline.format // "{emoji} {balance} ({name})"' "$R_CONFIG")
R_STATUS=$(echo "$R_FORMAT" | sed "s/{emoji}/$R_EMOJI/g; s/{balance}/$R_BALANCE/g; s/{name}/$R_NAME/g")
echo "$STATUS"
echo "$R_STATUS"
else
echo "$STATUS"
fi🌱 29 (contributor)
Run /respect-setup for an interactive wizard, or edit ~/.claude/respect/config.json directly. See config.example.json for all available options.
Customizable: tier names, thresholds, emoji, tip sizes, status line format.
Requires bats-core.
bats tests/test_tiers.bats tests/test_init.bats tests/test_tip.bats tests/test_correct.bats tests/test_session_start.bats tests/test_guards.bats81 tests covering wallet operations, tier calculations, guard execution, MCP tool matching, first-run onboarding, and global learnings.
See CONTRIBUTING.md.
MIT — see LICENSE.