feedback-loop-builder

Feedback Loop Builder

Make any Claude skill or agent self-improving. Retrofits a feedback loop that learns from usage and mistakes, getting smarter with every run — without ever changing the skill/agent's scope.

What It Does

• Analyzes your existing skill (SKILL.md) or agent (CLAUDE.md/AGENTS.md) to understand its structure and domain
• Injects three lifecycle phases (@BOOT, @REVIEW, @EVOLVE) that run automatically
• Learns patterns from user corrections, self-assessment, and repeated mistakes
• Asks before changing anything — every proposed improvement is shown to you with What/Why/Impact before being applied
• Validates changes using the skill-creator for skills, or an Opus agent review for agents
• Guards scope — the feedback loop only modifies feedback/patterns.md and feedback/run-log.jsonl, never the core behavior

How It Works

Every Run:
  @BOOT → Load learned patterns silently
  [... normal skill/agent work ...]
  @REVIEW → Self-assess: what worked, what didn't, user corrections
  @EVOLVE → Propose improvements → User approves/rejects → Patterns saved

After a few runs, your skill/agent accumulates domain-specific wisdom:

## Anti-Patterns (avoid these)
- Do NOT use pypdf for text extraction on scanned PDFs — use pdfplumber with OCR fallback

## Preferred Approaches
- When merging PDFs with different page sizes, normalize to first document's page size

## Edge Cases & Gotchas
- Password-protected PDFs: try empty string password before asking user

Prerequisites

1. Claude Code installed
2. An existing skill or agent to retrofit

Install

Option 1: From the Narai Marketplace (recommended)

/plugin marketplace add narailabs/narai-claude-plugins
/plugin install feedback-loop-builder@narai

Option 2: Install plugin directly from GitHub

/plugin install narailabs/claude-feedback-loop-builder

Option 3: Manual install

mkdir -p .claude/skills/feedback-loop-builder
curl -o .claude/skills/feedback-loop-builder/SKILL.md \
  https://raw.githubusercontent.com/narailabs/claude-feedback-loop-builder/main/skills/feedback-loop-builder/SKILL.md

Usage

# Retrofit a skill
/feedback-loop-builder path/to/my-skill

# Retrofit an agent
/feedback-loop-builder path/to/my-agent-project

The skill will:

1. Analyze the target and detect whether it's a skill or agent
2. Read the target's config to understand its domain
3. Create feedback/ infrastructure (patterns.md + run-log.jsonl)
4. Show you the tailored lifecycle sections before injecting
5. Ask for your approval before writing anything
6. Validate the result

Supports

Target Type	Config File	Detection
Skill	SKILL.md	Automatic
Agent	CLAUDE.md	Automatic
Agent	AGENTS.md	Automatic
Agent	agent.md	Automatic

License

MIT