Onboarding Claude Code Like a New Developer: Context, Skills, and the 17-Year Methodology

How to Onboard Claude Code Like a Senior Developer

Anthropic published a case study on April 28, 2026, showing how Brendan MacLean — a Claude Developer Ambassador at the University of Washington's MacCoss Lab — applied the same methodology he has used to onboard human developers for 17 years to Claude Code, with striking results.

The MacCoss Lab maintains Skyline, an open-source protein analysis platform with over 700,000 lines of C# code, a small team, and 200,000+ automated nightly tests. MacLean had tried Claude.ai for development work and found each session felt isolated — Claude had no persistent knowledge of the project structure, forcing repeated re-explanation of the same context.

The moment he switched to Claude Code and deliberately built out a context layer, everything changed.

The Three-Layer Context Architecture

MacLean's methodology organizes persistent knowledge into three distinct layers:

CLAUDE.md: The Lay of the Land

The CLAUDE.md file lives at the project root and is automatically injected into every Claude Code session. MacLean treats this file as an orientation document rather than a knowledge dump — it handles environment setup, explains the high-level project structure, and points to the documentation where the real expertise lives.

Critically, MacLean moved the AI context into its own separate repository (pwiz-ai), kept apart from the main codebase. This means the context applies across all branches and historical time points — not just the current branch state.

Skills: Persistent Domain Expertise

Skills are reusable knowledge modules that Claude Code applies on demand. MacLean's lab has built skills for skyline-development, version-control, and debugging — encoding the kind of institutional knowledge that would otherwise live only in the heads of long-tenured developers.

The "reference don't embed" principle guides skill design: CLAUDE.md points to documentation rather than inlining it, keeping files maintainable and avoiding context bloat.

MCP Integrations: Live Project Data

MacLean extended Claude's reach by connecting it to real project infrastructure via MCP. Claude Code now has access to nightly test results, crash reports, and customer support threads — the kind of real-time data that makes the difference between an assistant that describes problems and one that can actually investigate them.

Automated daily summaries pulled from nightly test infrastructure give Claude an up-to-date picture of the project's health at the start of every session.

Results That Made the Methodology Real

The impact became concrete when MacLean picked up a Files View panel that had sat unfinished for over a year after the original developer left the team. With Claude Code and a mature context layer, the feature shipped in two weeks — all final commits co-authored by Claude.

Other outcomes from the approach:

• CSS styling and feature additions that previously required designer involvement were handled independently by Claude Code
• Screenshot reproduction for over 2,000 tutorial images was fully automated
• Team members shifted from writing code to instructing Claude Code — spending their energy on direction rather than implementation

The Core Lesson: Context Is What Persists

MacLean's most important observation is also the one most developers skip: "Claude can't learn without you recording context. Don't expect magic."

To-do lists and plans Claude generates do not survive across sessions. The only thing that persists is what the developer deliberately maintains — the CLAUDE.md file, the skills library, the MCP connections. Teams that invest in this layer compound their gains over time. Teams that skip it find themselves re-explaining the same things session after session and never seeing the full leverage the tool can provide.

The post positions this context-maintenance discipline as the defining factor separating teams that see dramatic productivity improvements with Claude Code from teams that plateau after initial gains.

Applying the Methodology

For developers working on large or legacy codebases, the recommended starting point is running /init — the built-in command that has Claude explore the project and bootstrap an initial CLAUDE.md. From there, the methodology calls for iterative enrichment: adding skills as domain knowledge gets formalized, and connecting MCP servers as live data sources become valuable.

The MacCoss Lab story is particularly relevant for teams maintaining mature codebases where institutional knowledge is concentrated in a few people. Claude Code, given sufficient context, can act as a force multiplier for that knowledge — making it available to the whole team rather than locked in individual memory.