Operations Practice, Cost Governance, and Persistent Context

You're directing an operations and cost engagement for Sebastián Fuentes, VP of Infrastructure and Operations at Tienda Sur — a Chilean retail chain running on AWS with a 25-person platform team.

The discipline skills: an on-call rotation sized to a 5-engineer Tier-1 pool with compensation and surge protocol built in; severity definitions translated into the CFO's revenue language and the e-commerce VP's order-volume language; runbook standards on a cause/check/remediation triad; the operations-day shape across the four concurrent streams that fill 60-80% of the team's time; toil measurement as a recurring quarterly discipline with the 50% cap; a cost-governance framework with review cadences, anomaly detection, and an organisational-politics section; a vendor-management appendix on enterprise-agreement mechanics and the multi-cloud constraint; and a three-tier persistence pass over CLAUDE.md.

The AI-direction lesson is that operational practice and cost governance are organisational design problems, and AI does not see the organisation. It sizes the rotation to eight engineers when the team has five. It omits compensation and the Black Friday surge. It defaults to enterprise severity vocabulary the CFO does not speak. It writes a one-time toil audit and calls it a practice. It recommends a multi-year EDP without asking about the GCP evaluation. And it proposes every new artefact be added to CLAUDE.md as a Non-Negotiable, which dilutes the entries that already are. You design the practices AI does not, and decide what belongs in persistent context.

Your Role

You own the operations-practice document, the operations-day shape, the toil methodology, the cost-governance framework, the vendor-management appendix, the persistence-architecture pass over CLAUDE.md, and the team-adoption brief that asks the engineers to opt in. The Terraform footprint is small — a budget alarm, a cost-anomaly alarm, and an operations-KPI dashboard. The substance is the markdown and the judgment behind it.

Communication is Slack. Sebastián replies during Santiago business hours in conversational paragraphs with Spanish mixed in — warm by default, strategic underneath. Marco Reyes returns twice as edge-ratifier — once on rotation-with-compensation, once on toil-as-recurring-discipline — but does not propose the practice. You do.

What's New

Last time at Istanbul Technical University you designed an audit methodology, an import dependency graph, an OpenTofu evaluation, and an observability architecture across 12 research-group accounts.

This time two discipline columns culminate together. Operations and cost both reach their top character in the same engagement, because both are organisational design and both depend on the same governance bodies and the same persistent context. The audience is the 25 engineers who must adopt the practice voluntarily, with Sebastián as the channel and the CFO as the secondary stakeholder for the cost work. Communication switches from formal email back to Slack.

The hard part: a practice that doesn't account for the people who must follow it is documentation, not discipline. A rotation that needs eight engineers when you have five is unrunnable. A severity that lands in the wrong stakeholder's brain doesn't get acted on. A one-time toil audit is the wrong shape — toil is recurring, so the practice must be too. And a CLAUDE.md where every new entry arrives as a Non-Negotiable stops working as constraint architecture.

Tools

• Terraform — small footprint. Budget alarm in Unit 1, cost-anomaly alarm in Unit 6, operations-KPI dashboard in Unit 7.
• AWS Console and CLI — Cost Explorer as the data source for the governance cadence; Budgets as the alarm hook.
• AWS MCP server — carried-forward configuration only this engagement. mcp-config/aws-mcp.json ships with the allow-list, the archived classified-IP and GDPR-residency deny-list lineage, and the Project=tiendasur-ops scope filter, but no Unit invokes MCP at runtime — Unit 6's cost-anomaly demonstration uses aws ce get-cost-and-usage via the AWS CLI.
• CloudWatch and CloudWatch Logs — substrate for the dashboard and the routing path for the cost-anomaly alarm.
• CLAUDE.md — extended with four new entries across the three persistence tiers. Unit 8 is the architectural pass that decides which goes where.
• scripts/verify.sh — extended with one new dimension (check-ops-package-coverage.sh) asserting the operations package contains every required component before adoption.
• Mermaid and Markdown — the rotation calendar, the severity table with stakeholder-translation columns, and every framework document.

Materials

In the project zip:

• Sebastián's kickoff Slack message (sebastian-kickoff-slack.md) — three paragraphs at meta level, with the team-size, surge, and multi-cloud specifics held back for discovery.
• Tienda Sur organisational context (tiendasur-org-context.md) — the team role breakdown, the stakeholder roster, the cost-by-business-line split, and the Black Friday calendar.
• Team roster (team-roster.csv) — 25 anonymised rows with tenure, rotation-pool flag, timezone, and language preference.
• CFO letter context (cfo-letter-context.md) — the CFO's standing surprise-bill concern and the CLP authority thresholds.
• CLAUDE.md — engagement governance carried forward from Istanbul with four TODO markers where the new entries will land.
• Terraform starters (terraform/budget/, terraform/cost-anomaly/, terraform/operations-kpi-dashboard/) — the budget module is wired; the other two are scaffolds you fill in.
• Script starters in scripts/ — a TODO-body shell for the operations-package-coverage check, alongside the verify harness from Istanbul.
• A framework/ directory — empty, where every framework document will live.