You're running a cost review and first MCP activation for Craig Dunbar, Head of Cloud Operations at Ironstone Digital -- a mining technology company in Perth spending $180K a month on AWS across three accounts and unable to tell growth from waste.
The discipline skills: a budget alarm provisioned as the first resource in the account; a tag taxonomy and allocation strategy that turns a three-account tag mess into attributable spend; a systematic waste audit ranking the top fifteen items by annual AUD with a cost-of-change tier against the Flexera 29% benchmark; a Reserved Instance vs Savings Plan evaluation against actual usage; Cost Anomaly Detection provisioned via Terraform with two severity subscriptions; an AWS MCP server configured read-only with a service allowlist and trust-calibration probes to verify it before you rely on it.
The AI-direction lesson is that two things shift at once. AI stops generating about AWS and starts querying AWS, which changes the category of question it can answer and changes what verification has to look for. A count from MCP sounds authoritative but may silently include terminated instances, misread aggregation, or guess when the service is out of scope. At the same time, project memory becomes infrastructure: a CLAUDE.md that encodes the three-account topology, the predecessor RI history, and the "what not to touch" rules saves every future session the cold-start cost. You're practising both first-connection trust calibration and knowledge encoding that survives the session.
Your Role
Craig has read-only access for you on all three accounts and a pragmatic tolerance for waste that he's done tolerating. You direct Claude through the configuration and the analysis, verify output against Cost Explorer and the Console, and produce the artefacts Craig's team can re-run next quarter. The register stays collaborative -- Craig decides which waste is worth fixing, you surface the trade-offs with dollar amounts and cost-of-change tiers attached.
What's New
Last time in Bogota you built zero-downtime deployment and severity-based alert routing for Valentina -- blue/green with tested rollback, Kubernetes-vs-ECS on evidence, canary with evaluation criteria named up front, PagerDuty severity tiers wired to distinct SNS topics. The work was deployment safety plus routing design.
This time three columns advance together: AI Tool Connectivity activates from dormant to Mid, Cost & Capacity steps up to Mid-high for systematic rather than per-apply cost work, and AI Context reaches Late as project memory graduates from per-session to durable infrastructure. First contact is a plain email with dry Australian directness rather than a four-sentence Slack burst. Discovery is full and the hidden constraints Craig doesn't volunteer -- idle GPU drift, the mismatched predecessor RIs, the cross-AZ data transfer bill -- surface only if you ask about the right categories.
The hard part: MCP makes AI sound authoritative because the data is real, and that is exactly when miscounts and silent substitutions do the most damage. A confident answer from live infrastructure needs a different verification habit than a generated Terraform file did. You design the probes before you trust the connection.
Tools
- Claude Code -- drafts the Terraform and analysis; queries live AWS via MCP from Unit 3 onwards.
- AWS MCP server -- new: community server, read-only IAM, service allowlist, single region. Unit 3 walks through configuration.
- CLAUDE.md as authored artefact -- new: the track's first project-scoped memory file. Written in Unit 1 from discovery, extended across Units 3, 5, and 6.
- AWS Budgets, Cost Anomaly Detection, SNS -- the Terraform-provisioned cost infrastructure: a budget alarm with four notifications and a DIMENSIONAL anomaly monitor with two severity subscriptions.
- AWS Cost Explorer -- 90-day CSV export as the offline attribution dataset before MCP activates.
- Terraform CLI, AWS CLI v2, checkov, tflint, VS Code, Git/GitHub -- carrying forward. Three new commit prefixes:
cost:,tag:,mcp:.
Materials
In the project folder:
- Craig's first-contact email -- the short, plain-English message that triggers discovery.
- Account topology reference -- the three-account shape, synthetic IDs, tagging state, and "what not to touch" rules.
- CLAUDE.md, context-brief, tag-taxonomy, cost-allocation, waste-audit, handover, and README templates -- skeletons with TODO markers you fill as each unit surfaces the content.
- Cost Explorer 90-day export -- the offline spend dataset for the Unit 2 attribution work.
- AWS MCP reference, configuration template, trust-calibration probes -- the architecture, the scope discipline, and the probe pattern for Unit 3.
- Flexera benchmark, cost-of-change framework, RI/SP decision framework, anomaly detection reference, severity-tier examples -- the calibration and evaluation references for Units 4 and 5.
No filled CLAUDE.md, no tag taxonomy content, no waste audit numbers, no RI/SP recommendation, no anomaly thresholds -- you produce all of them.