I don't write much code anymore. I orchestrate. Here's the actual setup.
The mental model
Claude Code is a team, not a tool. The main conversation is the orchestrator — it routes, decides, and synthesizes. Specialized agents do the work in isolated contexts. MCPs handle service calls. Skills encode workflows.
If I'm reading files or writing Python, something's wrong with my routing.
The four-tier hierarchy
| Tier | What lives here | Why |
|---|---|---|
| Main context | Architecture decisions, skill routing, synthesis | Preserves "session state" across interactions |
| Subagents | Implementation, exploration, verification | Isolates token usage, parallelizes work |
| MCPs | Service calls (Supabase, Sentry, Vercel, PostHog) | Structured output beats parsing CLI stdout |
| Skills | Packaged workflows (ship, review, investigate, qa) | Forces discipline where I'd otherwise cut corners |
What this looks like in practice
"Ship the new booking flow" isn't me writing code. It's:
/shipskill routes me through test → review → commit → PRsecurity-revieweragent audits the diff in parallelmcp__plugin_supabase_supabase__apply_migrationruns the schema changemcp__plugin_vercel_vercel__triggers the preview deploy- I read the agent reports, approve or iterate
- Main context writes a handoff note to
tasks/session-state.mdfor the next session
Total time from "ship it" to merged PR: roughly 20 minutes for a typical feature. Most of that is me reading agent output.
The rule that actually matters
Delegate before you explore. The default trap is to start reading files to "get context" before routing. By the time you're oriented, you've burned 40% of your session on grep output that a subagent could have produced in parallel.
If a task will touch 2+ files, spawn an agent. If it needs deep codebase exploration, spawn an Explore agent. You write the prompt, it does the reading, you read the summary.
What I don't use Claude Code for
- Naming variables. I name things; AI picks boring names.
- Product decisions. AI has no skin in the game.
- Writing this post. I write the ideas, AI edits for tightness.
The skill ceiling of this workflow isn't "prompt better." It's "understand your own product better, so your taste scales." AI is leverage on judgment, not a substitute for it.