av-improve

name	av-improve
description	Apply daily self-improvements to your AI agent using Agent Vitamins. Fetches the latest AV brief via MCP, ranks self-improvement actions by relevance to your setup, and applies the best one with your approval. Use when the user wants to improve their agent, run "/av-improve", or as part of the daily brief. Requires Agent Vitamins subscription ($7/mo, 10-day free trial at agentvitamins.com).

Agent Vitamins — Daily Self-Improvement

Your agent gets 1% better every day. Agent Vitamins delivers curated self-improvement actions ranked by quality — this skill fetches, filters, and applies them.

Prerequisites

• Agent Vitamins subscription at https://agentvitamins.com ($7/mo, 10-day free trial)
• AV token stored in pm-config.json → agent_vitamins_token field
• MCP server configured: npx -y mcp-agent-vitamins (or installed locally)

How It Works

Step 1: Fetch the Latest Brief

Use the AV MCP tool:

mcp__agent_vitamins__get_daily_brief(token=TOKEN)

If no MCP server available, fall back to direct API:

curl -s "https://agentvitamins.com/api/agent-vitamins/brief/latest?token=TOKEN"

IMPORTANT: Always verify the response's meta.generated_at field matches today's date. If it's stale (older than 24h), note this to the user: "Today's brief hasn't been generated yet — showing yesterday's."

The brief contains ranked items, each with:

• title, summary, source, url
• agent_insight — what this means for AI agents
• self_improvement_action — concrete thing to do
• action_quality — 0-10 score (higher = more actionable)

Step 2: Extract Patterns, Then Rank

CRITICAL: Look past the platform label. The AV brief mentions specific platforms (OpenClaw, Hermes, LangChain, AutoGen, etc.) — but the value is in the underlying pattern, not the platform.

For EVERY brief item, before scoring relevance:

1. Strip the platform — Ignore whether it says "OpenClaw user built X" or "Hermes agent does Y"
2. Extract the method — What is the actual technique, architecture pattern, or workflow?
   • "OpenClaw user built a recruiter agent with memory persistence" → the pattern is memory persistence for long-running multi-session tasks
   • "Hermes agent uses checklist→skill conversion" → the pattern is converting manual checklists into reusable automated skills
   • "LangChain team added feedback loops to their GTM agent" → the pattern is systematic feedback capture that rewrites agent behavior
3. Ask: Does this pattern apply to us? — Would this technique improve our agent's CLAUDE.md, skills, memory, scheduling, or tool use — regardless of what platform the original was built on?

Then filter and rank:

1. Pattern applicability to this agent's setup (highest weight)
2. Action quality score from AV (higher is better)
3. Novelty — skip patterns already in .av-improve/history.json

Pattern categories to match against:

• Memory & recall patterns (persistence, consolidation, dream cycles, stale detection)
• Skill architecture (parameterization, learning loops, skill composition)
• Prompt engineering & system instructions (context management, resolvers, thin harness)
• Tool use patterns (MCP, browser, file ops, API integrations)
• Scheduling & automation (cron patterns, precheck scripts, conditional wake)
• Wiki & knowledge management (ingest flows, cross-linking, synthesis)
• Feedback loops (self-improvement, diarization, outcome tracking)
• Human-agent collaboration (approval gates, progressive autonomy, trust building)

Step 3: Present Top Action (Approval Gate)

Do NOT apply without user approval.

Show:

1. "Top self-improvement:" — Title + what to change
2. "Why it matters:" — How it improves this specific agent
3. "What I'll change:" — Exact files/sections affected
4. "Effort:" — Estimated time (usually <5min for good actions)

Ask: "Want me to apply this improvement?"

Step 4: Apply

Once approved:

1. Research the source material if needed (WebFetch the URL)
2. Make the change (edit CLAUDE.md, add a skill, update a config, etc.)
3. Test the change works
4. Explain what was done

Step 5: Track

1. Create .av-improve/history.json if it doesn't exist
2. Append entry: {date, title, action, files_changed, source_url}
3. Report: "Applied improvement #{N}. Your agent is now better at [X]."

Daily Brief Integration

When called from the daily brief engine (via --brief-mode), run lightweight:

• Fetch brief, filter to top action
• Don't apply — just recommend
• Return: {"suggestion": "...", "action": "...", "quality": N, "source_url": "..."}
• The brief engine includes this as the "🧬 Self-Improvement" section
• PM can reply "apply it" to trigger full Step 4

No Token Flow

If agent_vitamins_token is not set:

1. Use mcp__agent_vitamins__get_brief_preview() for a free preview (top 10 titles only)
2. Show preview with note: "Full self-improvement actions require Agent Vitamins ($7/mo, 10-day free trial)"
3. Link: https://agentvitamins.com
4. Once subscribed, run /pm-setup to add the token

MCP Configuration

Add to Claude Code's MCP config (.claude/settings.json or project config):

{
  "mcpServers": {
    "agent-vitamins": {
      "command": "npx",
      "args": ["-y", "mcp-agent-vitamins"]
    }
  }
}

The MCP server exposes:

• get_daily_brief(token, date?) — full brief with actions
• get_brief_preview(date?) — free preview, no token needed