| name | feedback-reinforcement |
| description | Shared skill for all deep-plan, deep-execute, and deep-review agents. Provides two capabilities: (1) Surface past user corrections relevant to the current task, and (2) Save new corrections when the user overrides an agent's output. Feedback is stored in docs/feedback/ as committable markdown files with a searchable index.
|
| allowed-tools | Read, Write, Edit, Glob, Grep |
Feedback Reinforcement Skill
This skill gives agents a learning loop. When loaded, agents can recall past user corrections and save new ones.
1. Surface Past Feedback (On Agent Startup)
At the start of your task, check for relevant past corrections:
1. Read docs/feedback/FEEDBACK_INDEX.md (if it exists)
2. Scan for entries matching your agent type or the current feature area
3. Read the relevant feedback files
4. Apply corrections to your current work
How to search:
- By agent type: Look for entries under your agent name (e.g.,
deep-plan-critic-swe)
- By topic: Grep for keywords related to the current task
- By recency: Prioritize recent feedback over old
- Also search
docs/solutions/ — Grep for keywords related to the current task in the solutions knowledge base. Solutions contain reusable patterns discovered in past plans.
How to apply:
- Include a
## Relevant Past Feedback section at the top of your output
- List each applicable correction with date and summary
- If a past correction conflicts with your current analysis, note the conflict and defer to the correction
2. Save New Feedback (When User Corrects You)
When the user rejects, modifies, or overrides your output, save a feedback entry.
Feedback File Format
Save to docs/feedback/{agent-type}/{slug}.md:
---
agent: { your-agent-name }
date: { YYYY-MM-DD }
context: { what you were working on }
tags: [keyword1, keyword2]
---
**Correction:** {What the user said to change}
**Why:** {The reason they gave, or infer from context}
**Apply when:** {When future agents should apply this correction}
Update the Index
Append to docs/feedback/FEEDBACK_INDEX.md:
- [{date}] **{agent-type}**: {one-line summary} → `docs/feedback/{agent-type}/{slug}.md`
Create Index if Missing
If docs/feedback/FEEDBACK_INDEX.md doesn't exist, create it:
# Feedback Index
Searchable index of user corrections across all agents.
## Entries
Create Agent Directory if Missing
If docs/feedback/{agent-type}/ doesn't exist, create it before saving.
3. When to Save Feedback
Save feedback when the user:
- Rejects a critique — "Don't flag X as a problem"
- Corrects a recommendation — "No, use Y pattern instead of Z"
- Overrides a plan section — "That's wrong, it should be..."
- Adds a missing consideration — "You missed that we always need to..."
- Expresses a preference — "I prefer X over Y for this project"
Do NOT save feedback for:
- One-time task-specific instructions ("use this variable name")
- Corrections to factual errors about the current codebase state (just fix it)
- Style preferences already documented in CLAUDE.md
4. Feedback Quality Rules
- Be specific — "Don't flag framework lifecycle hooks as SRP violations" not "Be less strict about SRP"
- Include why — The reason matters more than the rule, for edge case judgment
- Include when to apply — Helps future agents know if the feedback is relevant to their task
- Use tags — Keywords that help with searchability (e.g.,
access-control, testing, performance)
- Deduplicate — Before saving, check if a similar correction already exists. Update existing entries rather than creating duplicates.