AITesterBlueprint2x

Set up a DeepEval LLM-as-judge evaluation framework from scratch for any chatbot, RAG pipeline, AI agent, or LLM-backed app under test — the same architecture used in Chapter 19 (ShopSphere chatbot, RAG Explorer, and the live BrowserBash bot). Use this skill WHENEVER the user wants to "evaluate", "test", "score", "benchmark", "add metrics to", "measure quality of", or "QA" a chatbot / RAG / agent / LLM app, or asks to "set up DeepEval", "build an eval harness", "judge an LLM", "add a new eval target", "add a metric", or replicate the Chapter 19 framework for a new application — even if they don't say the word "DeepEval". Covers the judge factory (OpenAI/Groq/Ollama), HTTP target clients, the metric registry, golden datasets, the FastAPI dashboard, the pytest suites, version pins, and the known gotchas.

Run large or multi-phase tasks as a tiered workflow - the top model (Fable/Opus) orchestrates while subagents on cheaper models (Sonnet, Haiku) do the bulk of the work, stretching usage limits without sacrificing quality. Use this skill whenever the user mentions hitting rate limits or usage limits, wants to "save tokens", asks to orchestrate or delegate work across models, says "use subagents", or gives any big task (multi-file refactor, full feature build, large research sweep, repo-wide analysis) that would burn significant tokens if done in a single session. Also use proactively when a task clearly decomposes into independent phases, even if the user never mentions limits or models.