How I build software with AI-assisted development.

I use Claude Code with structured workflows that turn ideas into production code. The methodology is systematic: brainstorm, plan, implement with parallel agents, review, and verify.

User

Intent / Requirements

↕ Brainstorming

Design Spec

↕ Writing Plans

Implementation Plan

↕ Subagent-Driven Development

Parallel Agents

Isolated worktrees

↕ Code Review Checkpoints

Verified, Tested Code

Planning & Architecture

  • Structured Brainstorming

    Every project starts with collaborative design exploration. I work through requirements, constraints, and trade-offs before writing a single line of code. The output is a design spec that captures architecture decisions and their rationale.

  • Design Specs

    Written specifications document what we're building and why. They go through automated review loops to catch gaps and ambiguities before implementation begins.

  • Implementation Plans

    Specs become step-by-step plans with exact file paths, code patterns, and test strategies. Each task is small enough to implement in a single focused session.

Multi-Agent Orchestration

  • Parallel Agents

    Independent tasks run in isolated git worktrees with dedicated agents. Each agent gets precisely crafted context — just what it needs, nothing more.

  • Context Isolation

    Fresh agents per task prevent context pollution. The controller coordinates, dispatches, and reviews. Agents implement, test, and report.

  • Two-Stage Review

    Every task goes through spec compliance review first, then code quality review. Issues get fixed and re-reviewed in a loop until both pass.

Context Management

  • CLAUDE.md Files

    Project-specific instructions that persist across sessions. Architecture decisions, coding conventions, and gotchas are captured where they can inform every interaction.

  • Memory Systems

    Structured knowledge about projects, preferences, and decisions that carries forward. Not just what happened, but why it matters for future work.

  • Scoped Instructions

    Each project and sub-project gets its own context. Agents working on the kernel driver don't get confused by web frontend conventions.

Code Review & Verification

  • Automated Checkpoints

    Review gates between tasks catch issues early. Spec compliance ensures nothing is over-built or under-built. Code quality review ensures the implementation is clean.

  • Test-Driven Development

    Tests are written before implementation code. The red-green-refactor cycle keeps the codebase honest.

  • Verification Before Claims

    Nothing is declared 'done' without running the actual verification commands and confirming the output. Evidence before assertions.