/learn:feedback

Category: Learning Purpose: Manually submit feedback to the learning system

---

Usage

/learn:feedback                           # Interactive feedback
/learn:feedback --type correction         # Report a correction
/learn:feedback --type success            # Report what worked well
/learn:feedback --type agent-issue        # Report agent selection issue
/learn:feedback --type workflow-issue     # Report workflow issue

Script: ./scripts/learn/submit-feedback.sh

---

CRITICAL: Claude MUST Execute

When user runs /learn:feedback, Claude MUST actually submit to Supabase.

DO NOT just show documentation. DO run the curl command to store feedback.

---

Execution Steps

When user runs /learn:feedback, Claude should:

Step 1: Load Environment & Check

# Source .envrc first
source .envrc 2>/dev/null || source .claude/.envrc 2>/dev/null || true

# Verify
if [ -z "$SUPABASE_URL" ] || [ -z "$SUPABASE_SECRET_KEY" ]; then
  echo "❌ Learning system not configured"
  echo "Run: source .envrc"
  exit 1
fi

Step 2: Gather Feedback (Interactive)

Ask the user:

## 📝 Submit Feedback

**What type of feedback?**

1. ✅ **Success** - Something worked really well
2. ❌ **Correction** - I had to fix/redo something
3. 🤖 **Agent Issue** - Wrong agent was selected
4. 🔄 **Workflow Issue** - Problem with workflow phases
5. 💡 **Suggestion** - General improvement idea

**Select (1-5):**

Step 3: Collect Details Based on Type

For Success (type=success)

**What worked well?**
- Agent used: [auto-detect or ask]
- Task type: [e.g., "React component", "API endpoint"]
- Why it worked: [user input]

**Rate effectiveness (1-5):**

For Correction (type=correction)

**What needed correction?**
- What was the issue: [user input]
- What did you change: [user input]
- File(s) affected: [auto-detect recent edits or ask]
- Agent used: [auto-detect or ask]

**Severity (1-5):** 1=minor tweak, 5=complete redo

For Agent Issue (type=agent-issue)

**Agent Selection Problem**
- Task description: [user input]
- Agent selected: [auto-detect or ask]
- Better agent would be: [user input or suggest from list]
- Why: [user input]

For Workflow Issue (type=workflow-issue)

**Workflow Problem**
- Phase where issue occurred: [1-9]
- What happened: [user input]
- Expected behavior: [user input]

For Suggestion (type=suggestion)

**Improvement Suggestion**
- Category: [agent/workflow/rule/skill/other]
- Suggestion: [user input]
- Expected benefit: [user input]

Step 4: Submit to Supabase

# Determine feedback_type based on user selection
# success → approval
# correction → correction
# agent-issue → agent_override
# workflow-issue → rejection
# suggestion → suggestion

curl -X POST "${SUPABASE_URL}/rest/v1/af_feedback" \
  -H "apikey: ${SUPABASE_SECRET_KEY}" \
  -H "Authorization: Bearer ${SUPABASE_SECRET_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "feedback_type": "'"$feedback_type"'",
    "context": {
      "task_type": "'"$task_type"'",
      "agent_used": "'"$agent_used"'",
      "details": "'"$details"'"
    },
    "user_comment": "'"$user_comment"'",
    "severity": '"$severity"',
    "session_id": "'"$SESSION_ID"'"
  }'

Step 5: Confirm Submission

✅ Feedback submitted!

**Type:** Correction
**Severity:** 3/5
**Agent:** react-expert
**Details:** Component state management approach was overcomplicated

💡 Run `/learn:analyze` to see how your feedback improves the system.

---

Quick Feedback (Non-Interactive)

For fast feedback submission:

# Quick success
/learn:feedback --type success --message "TDD approach worked great for API"

# Quick correction
/learn:feedback --type correction --severity 3 --message "Had to simplify state logic"

# Quick agent issue
/learn:feedback --type agent-issue --agent react-expert --better nodejs-expert --message "Was backend task"

---

Feedback Types Reference

Type	feedback_type	When to Use
Success	approval	Agent/approach worked well
Correction	correction	You modified AI output
Agent Issue	agent_override	Wrong agent selected
Workflow Issue	rejection	Phase/workflow problem
Suggestion	suggestion	General improvement idea

---

Examples

Example 1: Report Successful Approach

User: /learn:feedback

Claude:
## 📝 Submit Feedback

What type of feedback?
1. ✅ Success
2. ❌ Correction
3. 🤖 Agent Issue
4. 🔄 Workflow Issue
5. 💡 Suggestion

User: 1

Claude: What worked well?

User: The react-expert agent's approach to form validation was perfect

Claude: Rate effectiveness (1-5):

User: 5

Claude:
✅ Feedback submitted!

Type: Success (approval)
Rating: 5/5
Agent: react-expert
Context: Form validation approach

💡 This positive feedback helps identify successful patterns!

Example 2: Report Correction

User: /learn:feedback --type correction

Claude: What needed correction?

User: The API error handling was too verbose, had to simplify

Claude: Severity (1=minor, 5=complete redo):

User: 2

Claude:
✅ Feedback submitted!

Type: Correction
Severity: 2/5
Details: API error handling too verbose, simplified

💡 This helps identify patterns to avoid.

Example 3: Quick Feedback

User: /learn:feedback --type success --message "Vue composition API pattern was exactly right"

Claude:
✅ Feedback submitted!

Type: Success
Message: Vue composition API pattern was exactly right
Agent: vue-expert (auto-detected)

💡 Quick feedback recorded!

---

Auto-Detection

The command will try to auto-detect:

• Current agent from session state
• Recent files from git status
• Task type from conversation context
• Session ID from environment

This reduces manual input needed.

---

Related Commands

• /learn:status - Check feedback count and system status
• /learn:analyze - Analyze feedback patterns
• /learn:apply - Apply improvements based on feedback

---

Version: 1.0.0