A developer's claude Code setup (11 commands, 3 agents) + shared Meta

A developer published a detailed Claude Code workflow on Substack that distills months of iteration into a reproducible setup: 11 custom slash commands, three coordinated agents, and a shared "design third brain" borrowed from Meta's internal design practice. The configuration is aimed at software engineers who want Claude Code to behave less like an autocomplete tool and more like a structured collaborator that remembers context, delegates subtasks, and follows house rules across a project's lifetime.

This piece breaks down what the setup ships, what it improves over a vanilla Claude Code install, and how a builder can replicate it today. It is most relevant to developers already running Claude Code who feel friction with long prompts, lost context, or repetitive scaffolding across sessions.

What shipped in A developer's claude Code setup (11 commands

The published setup centres on eleven slash commands that cover the most repetitive parts of a coding session: project bootstrapping, test generation, refactor planning, PR description drafting, and context reset. Each command is a plain markdown file stored in a .claude/commands/ directory at the project root, making the whole library version-controllable and shareable with a team via a single git clone.

Alongside the command library, the author wired three agents into the workflow. One agent handles planning and breaks down a feature request into discrete tasks. A second executes those tasks in focused, single-concern prompts. A third acts as a reviewer that checks the output against the project's stated constraints before the developer ever reads the diff. The three-agent loop is triggered by a single top-level slash command rather than three manual invocations.

What improved in A developer's claude Code setup (11 commands

The most structurally significant addition is the "design third brain" pattern, described as a shared memory layer that persists decisions — naming conventions, API shape choices, component hierarchy rules — across sessions. Without it, each Claude Code session starts cold, and the developer re-explains project-specific constraints through long system prompts. The third brain externalises that context into a living markdown document that every command references automatically.

The before/after contrast the author draws is primarily about prompt overhead and consistency. The table below summarises the qualitative shift described in the piece; no benchmark numbers were published:

What you gain from A developer's claude Code setup (11 commands

The primary benefit is session coherence. Because the design decisions live in a versioned markdown file rather than in the developer's memory or a chat transcript, a collaborator joining the project — human or agent — can pick up the same constraints without an onboarding conversation. The slash commands act as a shared vocabulary: anyone on the team types /plan-feature and gets the same structured output rather than a free-form response shaped by whoever wrote the prompt that day.

The three-agent architecture also changes how errors surface. Because a dedicated reviewer agent runs before the developer reviews output, trivial mistakes — wrong variable names, missing error handling, broken imports — are caught inside the pipeline. Developers who have used the setup report spending less time on the first-pass review cycle and more time on the architectural decisions the agents flag as ambiguous.

How to get A developer's claude Code setup (11 commands

To start, make sure Claude Code is current. Run the vendor update command in your terminal:

claude update

Once updated, create the commands directory at your project root and add your first slash command as a markdown file:

mkdir -p .claude/commands
touch .claude/commands/plan-feature.md

To launch Claude Code pinned to a specific model for a session:

claude --model claude-sonnet-4-5

To switch models inside a running session, use the in-app command:

/model claude-sonnet-4-5

To set a default model so every new session opens with the same configuration, add it to your Claude Code settings file rather than passing the flag each time — consult claude config for the exact key. The design third brain document should live at .claude/context.md and be referenced at the top of each slash command file so every command inherits the project's established decisions automatically.

What to watch after A developer's claude Code setup (11 commands

The three-agent pipeline currently runs sequentially, which means a planning agent, an execution agent, and a reviewer agent each consume a full context window in turn. For large codebases this can be slow, and the developer notes that parallelising the execution and review steps is the next obvious optimisation. Watch for updates to the Substack series that address concurrent agent execution and shared state between parallel runs.

The "design third brain" concept also raises a question that the current setup leaves open: how stale context is detected and pruned. A markdown file that accumulates decisions without a retirement process can mislead agents just as badly as no context at all. Builders adopting this pattern should establish a lightweight review cadence — perhaps tied to sprint boundaries — to audit and trim the context document before it grows beyond what a single agent can reliably use.