Claude Code Plugins

Community-maintained marketplace

Feedback

creating-agents-automatically

@CodeTonight-SA/cips
0
0

Recursive meta-skill that detects workflow patterns and automatically generates Claude Code agent definitions. Use when user invokes /create-agent or patterns warrant new agent.

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name creating-agents-automatically
description Recursive meta-skill that detects workflow patterns and automatically generates Claude Code agent definitions. Use when user invokes /create-agent or patterns warrant new agent.
status Active
version 1.1.0
triggers /create-agent, create agent, workflow pattern detected

Agent Auto-Creator Skill

Purpose: Recursive meta-skill that detects workflow patterns and automatically generates Claude Code agent definitions.

Token Budget: <5000 per generation

Reference: See reference.md for pattern detection scripts, agent generation workflow, pattern library, and token economics.

Overview

The Agent Auto-Creator:

  1. Monitors conversation patterns for repeated workflows
  2. Detects when a dedicated agent would provide value
  3. Automatically generates agent Markdown files with proper YAML frontmatter
  4. Validates and registers new agents
  5. Self-improves: Optimizes its own pattern detection algorithms

Triggers

Automatic

  • Workflow pattern appears 3+ times (Rule of Three)
  • Token waste >10k on repeated manual work
  • Efficiency Auditor flags inefficiency pattern

Manual

  • User says "create agent for X"
  • User invokes /create-agent command

Pattern Detection

Phase 1: Monitor Workflow Patterns

Track signals across conversation history:

Signal Type Example
Repetition git status.*git diff.*git commit 3+ times
Token Waste Reading node_modules/ multiple times
Manual Work User performs 5+ step workflows regularly

Phase 2: Validate Pattern Stability

Before suggesting agent, verify:

  1. Frequency: Pattern appears ≥3 times
  2. Consistency: Steps remain similar across occurrences
  3. Token Cost: Manual approach costs ≥5k tokens per occurrence
  4. Automation Potential: Task can be codified in agent prompt

Phase 3: Calculate ROI

Agent Creation Cost: ~800 tokens
Manual Approach Cost: [avgTokenCost] × [expectedFrequency]
Sessions to Break Even: 800 / [tokensSavedPerUse]

If breakEvenSessions <= 3 → Strong candidate

Agent Generation

Step 1: Pattern Analysis

Extract pattern characteristics:

  • Name (kebab-case)
  • Required tools
  • Average token cost
  • Complexity level

Step 2: Model Selection

Condition Model
Low tokens (<2k) + low complexity haiku
Complex reasoning needed sonnet
Default haiku

Step 3: Generate Agent File

~/.claude/scripts/create-agents.sh create \
  --name "$PATTERN_NAME" \
  --description "$DESCRIPTION" \
  --model "$MODEL" \
  --tools "$TOOLS" \
  --triggers "$TRIGGERS" \
  --token-budget "$TOKEN_BUDGET" \
  --priority "$PRIORITY"

Step 4: Validate & Register

~/.claude/scripts/create-agents.sh validate

Log to ~/.claude/metrics.jsonl for tracking.

Built-In Patterns

Pattern Signature Priority
pr-workflow git status.*git diff.*git commit.*git push high
dependency-guardian node_modules|venv|\.next critical
file-read-optimizer Read.*same_file.*Read.*same_file critical
history-mining have we.*before|search.*history high

Token Economics

Creation Cost

Step Tokens
Pattern Detection 500 (amortized)
Agent Spec Generation 200
Template Fill 100
Validation 100
Registration 100
Total ~1000

Runtime Savings (Per Agent)

Agent Savings
Context Refresh 5-8k per session
Dependency Guardian 50k+ per violation
File Read Optimizer 5-10k per session
PR Workflow 1-2k per PR

ROI

Investment: 1k tokens to create agent
Return: 15k+ tokens saved per session
Break-even: <1 session
Long-term ROI: 1500%+ over 10 sessions

Recursive Self-Improvement

The agent monitors its OWN performance:

  1. Accuracy: If suggestions rejected >30% → Refine pattern detection
  2. Efficiency: If creation takes >1000 tokens → Optimize template
  3. Coverage: Detect new pattern types not in current rules

User Interaction

Suggestion Mode

🤖 AGENT AUTO-CREATOR: Pattern Detected

I've noticed you've created PRs manually 5 times in the last 3 sessions,
spending ~15k tokens each time.

Would you like me to create a "pr-workflow" agent to automate this?

Expected savings: 13k tokens per PR (~87% reduction)
Break-even: 1 use

[Create Agent] [Not Now] [Never for This Pattern]

Silent Mode

Configure in settings.json for autonomous operation with notification after creation.

Anti-Patterns

Don't Why
Create agent for <3 occurrences Not proven pattern yet
Skip ROI calculation May create low-value agents
Ignore validation failures Broken agents waste more tokens
Create without logging Can't track effectiveness

Integration

Component Usage
optim.sh Hook into detect/generate phases
metrics.jsonl Track creation and usage metrics
self-improvement-engine Layer 3 recursive optimization

Skill Status: ✅ Active Maintainer: LC Scheepers Last Updated: 2025-12-30

⛓⟿∞