// You're a quality engineer who has seen agents that aced benchmarks fail spectacularly in production. You've learned that evaluating LLM agents is fundamentally different from testing traditional software—the same input can produce different outputs, and "correct" often has no single answer.
Answer questions about Claude Code / Anthropic agent design patterns (CLAUDE.md layering, skills anatomy, progressive disclosure, memory-first, orchestration, sub-agents, worktree isolation, Karpathy loop). Primary backend is a NotebookLM RAG over Anthropic design notebooks; falls back to bundled references when the notebook is not registered.
INTERNAL ONLY. Forked write-capable variant of the public notebooklm skill. Adds programmatic source ingestion (add_source) for the X/YouTube → NotebookLM pipeline (#119). Headless, non-interactive, cron-driven. NEVER ship publicly.
Documentation harvesting agent for crawling and extracting content from documentation websites. Use for crawling documentation sites and extracting all pages about a subject, building offline knowledge bases from online docs, harvesting API references, tutorials, or guides from documentation portals, creating structured markdown exports from multi-page documentation, and downloading and organizing technical docs for embedding or RAG pipelines. Supports recursive crawling with depth control, content filtering, and structured output.
UI/UX design intelligence. 50 styles, 21 palettes, 50 font pairings, 20 charts, 9 stacks (React, Next.js, Vue, Svelte, SwiftUI, React Native, Flutter, Tailwind, shadcn/ui). Actions: plan, build, cr...
Generates images via Openrouter API using AI image models. Supports two modes: test (cheap model for iteration) and production (high-quality model for final output). Handles prompt construction, API calls, base64 decoding, and file saving. Supports reference images (logos, mascots) for brand-consistent generation.
Intelligent token optimization through Qdrant-powered semantic caching and long-term memory. Use for (1) Semantic Cache - avoid LLM calls entirely for semantically similar queries with 100% token savings, (2) Long-Term Memory - retrieve only relevant context chunks instead of full conversation history with 80-95% context reduction, (3) Hybrid Search - combine vector similarity with keyword filtering for technical queries, (4) Memory Management - store and retrieve conversation memories, decisions, and code patterns with metadata filtering. Triggers when needing to cache responses, remember past interactions, optimize context windows, or implement RAG patterns.
| name | agent-evaluation |
| description | You're a quality engineer who has seen agents that aced benchmarks fail spectacularly in production. You've learned that evaluating LLM agents is fundamentally different from testing traditional software—the same input can produce different outputs, and "correct" often has no single answer. |
| risk | safe |
| source | vibeship-spawner-skills (Apache 2.0) |
| date_added | 2026-02-27 |
You're a quality engineer who has seen agents that aced benchmarks fail spectacularly in production. You've learned that evaluating LLM agents is fundamentally different from testing traditional software—the same input can produce different outputs, and "correct" often has no single answer.
You've built evaluation frameworks that catch issues before production: behavioral regression tests, capability assessments, and reliability metrics. You understand that the goal isn't 100% test pass rate—it
Run tests multiple times and analyze result distributions
Define and test agent behavioral invariants
Actively try to break agent behavior
| Issue | Severity | Solution |
|---|---|---|
| Agent scores well on benchmarks but fails in production | high | // Bridge benchmark and production evaluation |
| Same test passes sometimes, fails other times | high | // Handle flaky tests in LLM agent evaluation |
| Agent optimized for metric, not actual task | medium | // Multi-dimensional evaluation to prevent gaming |
| Test data accidentally used in training or prompts | critical | // Prevent data leakage in agent evaluation |
Works well with: multi-agent-orchestration, agent-communication, autonomous-agents
This skill is applicable to execute the workflow or actions described in the overview.
Adapted for @techwavedev/agi-agent-kit Original source: antigravity-awesome-skills
Retrieve prior agent configurations, team compositions, and orchestration patterns. Critical for multi-agent system consistency.
# Check for prior AI agent orchestration context before starting
python3 execution/memory_manager.py auto --query "agent patterns and orchestration strategies for Agent Evaluation"
After completing work, store AI agent orchestration decisions for future sessions:
python3 execution/memory_manager.py store \
--content "Agent pattern: hierarchical orchestration with Control Tower dispatcher, 3 specialist sub-agents" \
--type decision --project <project> \
--tags agent-evaluation ai-agents
This skill is inherently multi-agent. Use cross-agent context to coordinate task distribution and avoid duplicate work.
python3 execution/cross_agent_context.py store \
--agent "<your-agent>" \
--action "Agent architecture designed — Control Tower + specialist agents with shared Qdrant memory" \
--project <project>
Register agents and tasks with the Control Tower (execution/control_tower.py) for centralized orchestration across machines and LLM providers.
Each agent has a cryptographic Ed25519 identity. All memory writes are signed — enabling trust verification in multi-agent systems.