Claude is my primary tool. I don't hand it the thinking...I use it to get to a better decision faster, and I build the instruments that tell me whether it's actually helping: skills, guardrails and MCP servers around Claude Code, measured before they're trusted.
Read the map before reasoning
A markdown knowledge base the agent reads through an index before it answers, rather than working from memory of a previous session. It has its own MCP server with sensitivity rules enforced in code, an index builder, a leak checker and session hooks, all under test.
Keep secrets out of model context
A Claude Code plugin that masks declared secrets before they reach the model. Before sharing it I measured its false positives: 17 of 26 code shapes were wrongly masked. I fixed the design first, because a guard that corrupts source code is worse than no guard.
Encode a standard, then test it
Skills that carry a standard (a house writing voice, bounded research passes), plus an evaluation toolchain with an isolated runner and a local-model grader. A skill is a behaviour spec, so it gets regression-tested like one.
Measure the spend before optimizing it
I tested a context-compression proxy against my own session logs. It cleared my 30% gate (59% compressible), and I still didn't adopt it: tool output was at most 14% of context, so the real ceiling was about 5%. Output discipline in the project instructions did the same job with no infrastructure.
Claude Code
Primary tool for engineering work: multi-file changes, scripts, CI and infrastructure, running and fixing tests. Project instructions, skills and hooks set the rules, and the agent verifies its own work before I review it.
Claude chat and Cowork
Framing, tradeoff analysis, architecture decision records, and research passes filed into the knowledge base before any code exists. In regulated settings, also translating engineering work into compliance artifacts.
Opus, Sonnet or Haiku, per task
Model tier and thinking level are chosen per task, not set once. Judgment calls, code and bulk work each have a different best fit, and skills get re-checked when a new model ships, because a skill inherits the habits of the model it was written for.
Claude Certified Architect, Foundations
Passed September 2026. The exam covers agentic architecture and orchestration, tool design and MCP, Claude Code configuration, prompt engineering and structured output, and context management. Those five areas are what this page is built on.
Before generating anything, define the constraint space: architecture boundaries, team conventions, compliance requirements, timeline, and platform specifics like Bluetooth or background processing. The better the frame, the better the output.
Ask for three or four approaches with the tradeoffs written down. Optimize for decision quality, not speed alone.
Folder structure, layer boundaries, interface contracts and ownership come before any code. Coupling and testability get decided here, not discovered later.
Each change is a pull request that can be tested on its own before it merges. The PR is the gate, and tests ship with the change rather than after it.
Security, maintainability, test coverage, auditability and traceability. AI output is an input into an engineered system, not code to ship as-is.
At Inspire Medical Systems (a regulated medical device company) I was the first on my team to use GitHub Copilot in daily work. I was asked to show the engineering org how I was using it, then kept improving the team's approach: instruction files in the repo that the whole 8-person team used, and the team's architect pointed to them for the improvement in code quality.
Doing that inside a regulated environment shaped how I work now. PHI had to stay out of everything, the network was locked down (no MCP access at all), and API budgets could run out mid-week. AI adoption in a governed company is a policy problem as much as a capability problem. That work is where the five-phase workflow came from, and it carried straight over to Claude.
AI accelerates my work. It doesn't replace my judgment. I preload it with real constraints, shape the output space deliberately, and treat what comes back as a starting point for review. The failure mode I watch for isn't an obvious error...it's confident, well-formed output that's wrong, so anything that matters gets checked against a source before it ships.
If your team is adopting Claude, or AI generally, inside real constraints (regulated data, locked-down tooling, a budget someone has to defend), I'm glad to talk it through.