AI-Ops kit for a corporate operations team

In one paragraph

A mid-size operations team used to spend most of a working day stitching context together by hand — across a sales CRM, a support desk, an engineering tracker, an analytics warehouse, a payments processor, a BI surface, and the automation glue between them — just to answer a single business question. This kit replaces that scramble with one conversation. The operator asks in plain English; the right skill is selected automatically; a fleet of three specialist subagents investigates in parallel; an audit record is written without anyone having to remember to. Fifteen corporate skills reaching six systems, three subagents, built solo over about three weeks — and now shipped to the team as a finished, governed catalogue rather than a drawer of one-off scripts.

The problem in business terms

The team’s data lives in eight different places, each with its own login, its own query dialect, and its own quirks. Every real business question — why does this number differ across two dashboards? where did this customer’s refund get stuck? what’s actually on my plate before the sync call? — means reading two or three of those systems and reconciling the answers by hand. Before the kit, that reconciliation took multiple sessions and hours of elapsed time; operators without their own AI workflow simply escalated to engineering, and engineering became the bottleneck for routine questions.

What makes this different from “just install a few MCPs”

Most teams that bolt AI onto their operations stop at the MCP step: install a CRM connector, install a chat connector, hand Claude the credentials, hope for the best. That works for one user with simple questions; it doesn’t survive a second user or an audit. The kit’s value isn’t any single skill — it’s the fleet of three specialist subagents the main session delegates to. A multi-angle question doesn’t get answered in one stretched thread; Claude dispatches the warehouse-analyst, the CRM-agent, and the support-desk-agent to look in parallel, each loading only the references its corner of the world needs, each reporting back. That parallelism is what makes the kit feel like a competent analyst rather than a bot that takes forever and forgets what you asked — and crucially, none of the subagents grants new access; each inherits exactly the access the operator already has.

The two decisions that make it safe to hand to the whole team are quieter. Credentials never pass through the model: every skill reads its token from the OS keychain at execution time and uses it locally, so if a session transcript leaked tomorrow, no secret would leak with it — the single most under-discussed pattern in AI-ops today. And every skill carries a RACI block at the top of its instruction file by template, not by hope, so when a skill returns a wrong answer six months from now, accountability is in the file rather than in someone’s memory. [devil's-advocate]

A session in the room

Three everyday operator questions, run consecutively — each one touching a different corner of the business, each answered end to end. The agent investigates, decides, executes, and verifies; every scenario closes on a success line. The greeting pinned on the right stays put as you scroll; each scene below picks up the next task as its centerline crosses.

The catalogue

Each connector domain is a read skill and a write skill; the write skill always inherits the extra discipline shown in the table above. Three meta-skills sit above the connectors — one to scaffold new skills, one to onboard a new operator, one to refresh the dated snapshots inside other skills’ references.

The three subagents — warehouse-analyst, crm-agent, desk-agent — are the value layer. Each is a read-only specialist that loads only its own references, runs in its own context window so the main thread stays clean, and reports findings back to the orchestrator. None grants new access; each inherits the operator’s existing scope. Three more connectors — a payments processor, team chat, and a secondary accounting system — are scoped but deliberately out of this kit’s first version; adding one is now a mechanical pass through the meta-skill, not a design problem. The architecture is done.

Governance: the merge gate

Every artifact passes this checklist before it can merge. Lint enforces the mechanical items; reviewers enforce the rest.

1. Useful to at least two people — single-user helpers stay personal, outside the corporate kit.
2. A named owner — a RACI block at the top of the instruction file, no exceptions.
3. Template-compliant — required sections present, no TODO markers, plain-language triggers.
4. Safe defaults — read-only by default; no destructive operation runs without confirmation.
5. No secrets — verified by lint and the CI secret-detection job.
6. No name conflicts — the area-verb name is unique across the catalogue.
7. Documented failure behaviour — what the skill does when the API errors, auth expires, or input is malformed.

A lint rule that protects the team starts life as a warning and graduates to a hard error only after a few real skills have used it successfully in production. The rule engine evolves with the catalogue rather than running ahead of it — governance that earns its strictness instead of asserting it.

The audit trail, in two sentences

Every skill that touches an upstream system is wrapped in a

@telemetered(
  skill="warehouse-read",
  operation="query",
  outcome_map={0: "ok", 1: "error", 2: "auth_expired"},
  extras=lambda result, args: {"sql_preview": args.sql[:200]},
)
def cmd_query(args): ...

that captures what skill ran, against what system, with what parameters, with what outcome, and how long it took — and ships those records to a central log aggregator. Two operational properties make it trustworthy: the shipper is idempotent (a session that drops its connection mid-flight retries later and never ships a duplicate), and it is non-blocking (if the aggregator is unreachable, records queue locally and the operator’s session is unaffected).

For the business, that’s a complete record of who asked the AI tier what, against which system, when, and with what result — the foundation for trusting the kit at scale, and the starting point for any future compliance conversation.

The verdict from the room

The kit’s impact is best measured by what stops happening: operators no longer queue behind engineering for routine questions, and the answers they get arrive already reconciled and already recorded. An operator who uses it daily puts it more plainly than any metric — “there’s nothing to configure; it does everything it can on its own, and when it does need me it’s one simple step with clear instructions. It structures every answer cleanly — everything laid out in its place.” That is the whole design goal stated as a user experience: the machinery is invisible, the judgement is visible, and the operator is never asked to think about the plumbing.

~3 weeks. One builder. 15 skills, 3 subagents, 6 systems. A multi-system investigation that took a day now takes minutes — reconciled, recorded, and safe to hand to the whole team.

This case describes architecture and patterns. Specific vendors, hostnames, and the client itself are deliberately left abstract; what matters here is the shape of the system — one conversation over many systems, governed so a team can trust it — not the procurement choices behind it.