You're designing AI infrastructure for Noa Levi, Platform Engineering Lead at MediConnect, a Tel Aviv telehealth company where thirty developers keep shipping the same infrastructure mistakes despite a written guidelines doc nobody reads.
The discipline skills: an encoded-standards CLAUDE.md that tells AI the conventions, security policies, and compliance non-negotiables before the prompt arrives; a multi-dimensional verification system that separates security, compliance, and cost into distinct channels with distinct tools; and a delegated audit agent with read-only AWS access and a defined output shape.
The AI-direction lesson is that three things shift at once. Project memory stops being session context and becomes the place you encode standards -- the contrast experiment in Unit 1 makes this visible by running the same prompt cold and equipped. Verification stops being a checklist you run on each output and becomes a system you design across dimensions. And work you used to review yourself starts moving to a constrained agent whose scope, tool access, and output shape you decide on purpose.
Your Role
Noa is technically sharp and frameworks-fluent. She has named the problem in terms of context, dimensions, and scope before you arrive, so you're not teaching her concepts -- you're designing the architecture she has named the problem for. You direct Claude through the standards drafting, the verification design, and the agent configuration. You verify by running the system her developers will use, not by hand-checking each AI output. The register moves toward "AI as tool requiring active management": you decide what AI knows at session start, what verification looks like across the team, and what a delegated agent is allowed to do.
What's New
Last time at Ironstone you connected AWS via MCP read-only and wrote your first CLAUDE.md -- the project-specific facts about Craig's three accounts and the resources not to touch. That CLAUDE.md captured what's true about THIS project.
This time CLAUDE.md takes on its second job: capturing what's true about HOW WE WORK on this project. The format is the same; the content category is different. You'll see both layers and learn which goes where. Verification graduates from a per-output check to a designed system with three dimensions routed to three tools. And for the first time you configure a delegated agent whose scope, tool access, and output shape you decide before it runs.
The hard part: encoded standards reduce error rate but don't eliminate it. AI sometimes ignores its own context, especially when the file is too long or mixes priorities with preferences. The standards file has to be structured for AI to hold all of it -- and the audit agent has to catch what slipped through.
Tools
- Claude Code -- drafts the standards, the policy files, the verification scripts, and the agent configuration.
- CLAUDE.md as encoded standards -- new substance: conventions and policies, not project facts. Structure (non-negotiables / conventions / preferences) matters more than length.
- AGENTS.md -- new lightweight portability companion so a developer using Cursor or Copilot gets the same standards.
- checkov -- carrying forward; the security dimension.
- conftest + Open Policy Agent (Rego) -- new: the compliance dimension. HIPAA-derived rules checkov doesn't cover natively.
- Infracost -- carrying forward; the cost dimension.
- Claude Code subagents -- new: the delegated audit agent is a definition file under
.claude/agents/with explicit tool access, system prompt, and output shape. - AWS MCP server, AWS Console + CLI, Terraform, Git/GitHub, AWS Budgets -- carrying forward. The MCP discipline (read-only, allowlist) becomes the agent-scope discipline.
Materials
In the project folder:
- Noa's first-contact Slack message -- the short, framework-language opener that triggers full discovery.
- Noa's existing guidelines doc -- a deliberately disorganised 200-line real-world artefact: security policies mixed with naming preferences and aspirational comments. You reorganise and condense it.
- HIPAA controls reference -- the subset of HIPAA Security Rule controls relevant to AWS infrastructure, mapped to service configurations in plain English.
- Verification architecture reference -- the pattern (security / compliance / cost as separate channels), the three-tool starter set, and the question of "which tool for which check" left as your design.
- Subagent reference -- how
.claude/agents/works at the file-format level, and the relationship between system prompt, tool access, and output shape. - CLAUDE.md.starter and AGENTS.md.starter -- near-empty files with the section structure named and one example rule per section, so you see the structural discipline without being given the substance.
You produce the encoded standards, the policy files, the orchestrating verification script, and the agent definition. The exemplar CLAUDE.md is held back -- you see it later as a sanity check, not a template to copy.