You're consolidating operations for Lucía Hernández, VP of Platform Engineering at Conecta Móvil, a Mexican telecom serving 8 million subscribers, where 15 microservices are deployed by hand and the platform team spends 70% of its time on repetitive operational work.
The discipline skills: a toil inventory that classifies each repetitive task against the precise six-criterion definition, with a build-cost / maintenance-cost / time-saved triple on every automation candidate; a service dependency graph extracted from the codebase and the team's lived knowledge; a GitOps deployment under ArgoCD where sync waves enforce ordering; OpenTelemetry emitting metrics, logs, and traces through one library with sampling sized to a budget; dashboards and alerts declared in Terraform; and three MCP connections with permission scopes designed per channel.
The AI-direction lesson is that AI does not name what it leaves out. Asked to propose automations, it lists every repetitive task without applying the toil definition or naming build cost. Asked to deploy 15 services, it ships them in alphabetical order. Asked to instrument tracing, it captures every span. Asked to set up monitoring, it walks through the Datadog console UI. Asked to connect tools, it requests the broadest scope on every channel.
Your Role
Lucía is technically deep. She implemented distributed tracing at her previous company, references Google SRE on toil, and names her team's drag in business metrics. What she lacks is bandwidth — her team is too busy running operations to engineer their way out. The register opens with a 30-second voicemail and continues in Slack, with a weekly written status update in business-metric language. You direct Claude through inventory, dependency analysis, the ArgoCD bootstrap, the OpenTelemetry instrumentation, the monitoring-as-code, and the permission design. You verify by running the system, naming the cost, and catching the defaults AI did not name.
What's New
Last time at the Abu Dhabi Digital Authority you executed state surgery on a twelve-entity consolidation, authored sovereign-data policy in Rego, and provisioned a remote state backend. You leave that engagement with the multi-dimensional verification habit, the policy-as-code authoring loop, and a CLAUDE.md that grows with the work.
This time the operational stream is the substance. Toil stops being a feeling and becomes a six-criterion definition you apply task by task. Deployment stops being one app to staging and production and becomes 15 services with a real dependency graph, deployed by a controller pulling desired state from Git. Observability stops being a metric and an alert and becomes three pillars together — metrics, logs, traces — through one instrumentation library, with sampling sized to a monthly budget. AI tooling stops being one connection and becomes three, each with a permission scope you design rather than accept.
The hard part: AI's reflexes are productive in isolation. Parallel deployment is fast. Every span captured is complete. Every metric is comprehensive. Full admin is convenient. Each default is reasonable on its own and dangerous together.
Tools
- Terraform — now the substrate for monitoring-as-code (Datadog provider) and the ArgoCD bootstrap.
- OPA + conftest — extended with a deployment-dependency-graph linter you author.
- CLAUDE.md — extended with two new Conventions: the deployment-order rule and the per-MCP permission scope catalogue.
scripts/verify.sh— extended with the dependency-graph dimension.- AWS, Datadog, and GitHub MCP servers — three connections, each read-only and scoped per channel.
- Datadog — new, the unified observability backend.
- OpenTelemetry SDK and Collector — new, one instrumentation library for metrics, logs, and traces.
- ArgoCD — new, the GitOps controller; sync waves are the ordering primitive.
- Helm — new, the manifest format ArgoCD applies.
- AWS Console + CLI, Cost Explorer, kubectl, Claude Code, AWS Budgets — carrying forward.
Materials
In the project zip:
- Lucía's voicemail transcript (
voicemail-transcript.md) — 30 seconds, asks for a callback. - Microservice inventory (
services-inventory.csv) — 15 rows with name, language, runtime, repo path, and role. The dependency graph is not in the file. - Legacy deployment scripts (
legacy-deployment-scripts/) — three bash scripts capturing the manual procedure the team runs today. - Current monitoring snapshot (
monitoring-snapshot.md) — the existing stack with monthly cost ranges per component. - Helm chart starters (
helm-charts/) — three representative charts (API, worker, batch) with staging and production values. - ArgoCD bootstrap reference (
argocd-bootstrap-reference.md) — names the pattern; the wave-and-app-name design is yours. - OpenTelemetry collector reference (
otel-collector-reference.yaml) — a working collector for one service. Sampling rates per tier are yours. - CLAUDE.md — engagement facts in place and two TODO markers for the new Conventions.
You produce the toil inventory, the dependency graph, the ArgoCD applications and sync waves, the OpenTelemetry instrumentation and sampling, the Terraform-declared dashboards and alerts, the per-MCP scopes, the dependency-graph linter, and the handover roadmap.