Example · fictional scenario
A Claude skill foryour internalinfrastructure.
A worked example — deliberately invented, but realistic. This is what an AI tool would look like when it runs on your own infrastructure, not in a vendor's cloud.
Situation
A company of 15 people. The operational knowledge is scattered: half in the wiki, half in old Slack threads, plus a PostgreSQL database holding the inventory of systems. Whoever is on call at night spends longer finding the right information than the problem itself takes.
Approach
First the questions: which workflows really cost time? Which data must never leave your own server? What is already documented, and what only lives in people's heads?
From that, two building blocks. An MCP server with read-only access to the inventory database — only through fixed, approved queries, not open SQL. And a Claude skill that knows the team's runbook conventions: how to escalate, which checks apply before any change, where the limits are. Both run on your own infrastructure.
Architecture
Where the limits are.
One path, clearly bounded: the request runs through the skill to the MCP server, which only issues approved read queries against the database. Guardrails check every step, every action lands in the audit log. No open database access, no leak to the outside.
Result
The team asks in plain language: “Which systems hang off this switch?”, “What was the last maintenance step for the DB cluster?” — and gets an answer with a source, not a guess. Operating cost stays small because the routine goes to a cheap model and only the hard cases go to the expensive one.
Deliberately not built: no write access to the database, no automatic changes to systems, no data in someone else's cloud. The tool answers questions and suggests paths — deciding and acting stay with people.
Your infrastructure looks different. The way there is the same: understand first, build narrowly, then prove it. I build the tools for that myself, every day — lore and context-parachute are public.