This is Part 6 of a series on plan-based development with Claude Code. Today we explore how Claude Code becomes an integrated development partner.
From Chatbot to Development Partner
Claude Code isn’t a chatbot you occasionally ask questions. It’s an integrated development partner that understands your codebase, follows your patterns, and executes multi-step tasks.
The difference is profound:
- Chat: “How do I add a new route in Express?”
- Claude Code: “Add a new protected route for user preferences following our existing patterns in
@bts/api”
Custom Commands: Codified Workflows
We created custom commands that encode our team’s workflows:
.claude/├── commands/│ ├── research.md # /research - Research & planning│ ├── understand.md # /understand - Code comprehension│ ├── find.md # /find - Code discovery│ ├── test.md # /test - Testing workflow│ ├── review.md # /review - PR-style code review│ └── write.md # /write - Code generationThe /research Command
When planning a new feature:
/research add real-time notifications to the dashboardThis triggers a structured research process:
- Searches for existing notification patterns
- Reviews the current architecture
- Identifies affected packages
- Produces a structured implementation plan
The /review Command
Before committing:
/review stagedThis performs a comprehensive code review:
- Checks for bugs and security issues
- Validates TypeScript patterns
- Ensures consistency with repo conventions
- Suggests improvements
Skills: Context-Aware Assistance
Unlike commands that you invoke explicitly, skills activate automatically based on context:
## When to Activate- Questions about TypeScript errors- Code with type assertions or `any`- Discussions of validation or schemas
## Guidance- Prefer Zod schemas over manual type guards- Use branded types for IDs- Validate at system boundariesWhen Claude sees code with as any, it automatically applies the type-safety skill’s guidance.
The Plan Mode Workflow
For complex features, we use Claude Code’s plan mode:
- Enter plan mode: Triggered for multi-step implementations
- Explore codebase: Claude reads relevant files, understands patterns
- Write plan: Creates detailed implementation steps
- Review: Human reviews and approves the plan
- Execute: Claude implements following the approved plan
This mirrors our manual plan-based development but with AI assistance at every step.
Real Examples from Our Workflow
Example 1: Fixing a Build Issue
> bun run dev shows "Cannot find module '@bts/effect/services'"
Claude analyzes:- Checks turbo.json for task dependencies- Identifies race condition: dev starts before build completes- Proposes fix: Add "^build" to dev task dependencies- Implements the fix- Verifies with test runExample 2: Adding a New Package
> Create a new @bts/highlighter-router package following existing patterns
Claude:- Uses `turbo gen package` generator- Creates turbo.json with correct tags- Sets up tsconfig.json extending base- Adds to workspace catalog- Creates initial exports- Updates @bts/router to include new routerThe Amplification Effect
Claude Code doesn’t replace developer judgment—it amplifies it.
- Without Claude: Manually search 60 packages, copy patterns, adapt to context
- With Claude: Describe intent, review proposed solution, approve and refine
The cognitive load shifts from “how do I do this?” to “is this the right approach?”
Next up in Part 7: Putting it all together - the virtuous cycle and lessons learned from six months of this workflow.