FAQ

It means I design the workflow before I pick the model. I map human steps, decide where judgment is required, then assign Cursor for code, MCP servers for live APIs, n8n for triggers, and Python for batch jobs. The LLM is one station on an assembly line I can test, log, and swap — not the whole factory.

ChatGPT is great for thinking out loud, but production work needs boundaries. I use models inside pipelines with structured outputs, retrieval, tool permissions, and eval gates. When a workflow repeats, I automate it; when it needs taste or accountability, I keep a human in the loop.

If the steps are fixed and the data shape is stable, I reach for n8n or a Python script first. Agents earn their place when intent varies, context matters, or tool selection depends on the request. I have seen teams burn weeks agentifying a flow that a cron job could own.

An eval question set and a logging skeleton. Before prompts or UI, I write twenty questions stakeholders actually ask, plus failure cases I am afraid of. Those become regression tests, FAQ seeds, and the acceptance bar for shipping.

I time-box discovery, then ship a narrow vertical slice with metrics: latency, accuracy on the eval set, and human review rate. If we cannot measure it, we are still in demo land. Production AI systems need a definition of done like any other feature.

How much of the work is integration, not prompting. Wiring Gmail, CRM, GitHub, and vector stores through MCP took more calendar time than tuning system messages — but it is also what made outputs trustworthy and repeatable.

Model Context Protocol servers expose typed tools — search inbox, fetch ticket, create draft — that agents call at runtime. I build custom MCP layers over Gmail, GitHub, Supabase, and Stripe so the model never gets raw admin keys. It is the difference between a demo chat and an operator that can act safely.

When multiple agents or IDEs need the same capability, or when you want audit logs and permission scopes in one place. I wrapped our CRM lookups once in MCP and every workflow — support copilot, email triage, reporting — reused the same tool definitions.