prompt-engineer

Use ANY time an LLM-powered agent, chatbot, or AI assistant is being built, debugged, reviewed, planned, or migrated — the user almost never says 'prompt.' They say 'the agent keeps calling the wrong tool,' 'why is it ignoring my instructions,' 'fix this bug,' or just paste a log. TRIGGER when (1) any file in scope contains a system prompt, tool/function description, function-calling schema, tools.json, agents/* file, or imports an LLM SDK or agent framework (anthropic, openai, google.genai, AWS Bedrock, Vertex, Azure OpenAI, LangGraph/LangChain); (2) an agent loops, picks the wrong tool, hallucinates arguments, runs verbose despite low-verbosity settings, over- or under-escalates, ignores instructions, or shifts behavior after a model upgrade; (3) the user is thinking through or planning agent architecture — subagents, orchestrator/worker splits, state passing, multi-agent workflows — before code is written; (4) the user is comparing providers, migrating a prompt to a newer model, or asking 'is this good enough.' Covers Claude/Anthropic, GPT/OpenAI (incl. Azure), and Gemini/Google (incl. Vertex). Invoke proactively the moment you see prompt-shaped text or an LLM SDK call in scope — do not wait for the user to ask. SKIP only when no LLM agent is in the loop (pure SQL, infra, normal code with no model in scope, human prose editing).