← Back to Blog

20 Days, 20,000 Conversations, 1.2 Billion Tokens — What I Learned Using Claude Code

From treating Claude Code as "a smarter Cursor" in November to managing it like a reliable outsourced team today — this is a full retrospective after 20,000 conversations. Officially recognized as a top 1% Claude Code user by token consumption.

The Mindset Shift: From AI Assistant to AI Team

When I first started using Claude Code in November, I thought of it as a smarter Cursor — just more accurate at writing code.

Then I looked up from my screen. Through conversations with colleagues and observing the broader AI landscape, I realized Claude Code can no longer be defined as a "coding AI assistant." It's not a knowledgeable intern anymore. It's —

An experienced, cost-effective, reliable team.

This shift fundamentally changed how I use Claude Code. Instead of "write me a function," it became "I have a product requirement — take it from PRD to production."

Results: 3 Products in 20 Days

Claude Code now has the ability to independently ship a complete product. In 20 days, I built:

Product Stack Time Notes
AI Chat Bot Flutter 2 days Voice/text input, resume upload, AI interviewer
iOS Launcher App Native iOS 2 days Quick prototype validation
GroAsk Native macOS 2 weeks Claude Code dispatch panel

All three products were pure Vibe Coding — I didn't write a single line of code by hand. During the Opus 4.5 era, I still reviewed code occasionally. With Opus 4.6, I haven't looked at a single line.

GroAsk was born from my own frustration managing multiple Claude Code sessions — constantly switching terminals, forgetting which session was doing what, and not knowing when context was about to overflow. The multi-session management and context management techniques I'll discuss below are directly related to it.

Beyond Coding: What Else Can Claude Code Do?

Most people only use Claude Code for programming, but its capabilities extend far beyond:

  • Auto-posting to Twitter — Using browser MCP and Skills to compose and publish automatically
  • Writing articles & multi-platform distribution — Custom Skills that generate platform-adapted articles in one shot
  • Product analytics — Analyzing real operational data for retention, conversion, and channel performance
  • Deep market research — Competitive analysis, user pain point collection, market trend analysis

Claude Code is fundamentally a general-purpose task execution engine. Programming is just its strongest vertical.

My Product Iteration Workflow

Small features: One sentence to production

Describe the requirement → Claude Code handles it → Ship → Fix issues with one more sentence

Most daily needs work this way. "Change the menu bar icon from brain to lightning bolt" — one sentence, 30 seconds, done.

Major features: Structured process

Requirements → PRD → Market research validation → Technical spec → Execution plan → Build

In practice:

  1. Requirements: Personal pain points + Claude Code community feedback
  2. PRD: Describe raw requirements, let Claude Code produce a PRD, validate with market research
  3. Technical spec: Claude Code writes the spec, reviewed against market research
  4. Execution plan: Claude Code uses TDD + Subagents to break down tasks
  5. Development: Claude Code executes the plan

A 3,000-word PRD → 10,000-word technical spec → 40,000-word execution plan

Core Principle

Product decisions and technical decisions are the human's most important job. Fully delegating decisions to AI still isn't reliable for complex tasks — it looks like it worked, but the maintenance burden is nightmarish. Getting the plan right is everything.

Having AI implement one small requirement at a time is easier, more reliable, and more efficient than asking it to build everything at once.

Power User Techniques

Skills: The Most Powerful Feature

Core idea: Codify your personal workflows.

Skills are Claude Code's most underrated feature. They turn repetitive workflows into single commands.

Two essential parts of a Skill:

  • Prompt: Natural language for parts requiring model judgment
  • Scripts: Code for deterministic operations

Only use prompts for what requires model understanding. Everything else should be a script.

My commonly used Skills:

  • /dev-ship: Atomic commit + push
  • /dev-release: Build → sign → notarize → GitHub Release
  • /write-post: Generate platform-adapted distribution articles
  • /ops-status: Pull product metrics in one command

The benefit of codifying workflows into Skills: consistent quality every time. No more skipping steps because you're tired.

Subagents: Context Management for Complex Tasks

Subagents let you use multiple agents within a single session. Claude's built-in Task tool is a Subagent — the main session delegates work to internal sessions and only cares about inputs and outputs.

Use cases:

  • Parallel processing of independent subtasks (e.g., generating articles for 16 platforms simultaneously)
  • Context isolation (subtask intermediate states don't pollute the main session)
  • Complex task decomposition (main session orchestrates, sub-sessions execute)

MCP: Extending Claude Code's Capabilities

  • I personally use 8 MCPs, but strongly recommend fewer
  • MCPs consume significant base context — use parameters to only expose necessary interfaces
  • Prefer Rust/Go compiled MCPs over Node-based ones for significantly lower memory usage

Hooks: Your Safety Net

The most valuable hook right now: Make Claude Code back up anything before deleting it.

Hooks are event triggers that run your scripts before or after specific operations. I use them for safety checks: verifying no print() debug statements before commits, ensuring config files are synced, etc.

The Ultimate Debugging Technique

Have AI write logs locally, then debug based on those logs. This is the single most effective bug-solving approach.

When you think AI can't figure it out, try logs. The results will exceed your expectations. AI often can't see runtime state — logs turn the black box into a white box, and problems become immediately clear.

Long-Term Memory: Solving "Knows Today, Forgets Tomorrow"

Long-term memory is a common pain point. You tell Claude Code about your project architecture, coding standards, and preferences today — open a new session tomorrow and it's all gone.

My solution: Knowledge grouping.

  1. Maintain a layered Knowledge folder and instruct Claude Code to "store valuable information in knowledge"
  2. Layered CLAUDE.md management:
    • Root CLAUDE.md: Global project info
    • Subdirectory CLAUDE.md: Sub-project specific standards
    • Rules directory: Coding conventions, lessons learned
  3. Achieve local long-term memory that improves with use

The key is layering — don't cram everything into one file. Claude Code's context is limited. The more precise and structured the information, the better the model understands it.

Context Management: Keeping the Model Smart

Context management is the single biggest factor affecting Claude Code experience, and the most overlooked.

Strategy Details
Load MCP tools on demand Use parameters to minimize exposed interfaces
Layer CLAUDE.md files Avoid information overload
Context threshold: 200k Switch sessions beyond this
Avoid context compression New session >> compressed context
Real-time monitoring Know each session's context consumption

Don't let context exceed 200k (20% of the 1M window). The difference is like a colleague at 9 AM vs. one pulling an all-nighter — by midnight, their thinking is muddled. Always use the smartest model.

That last point — real-time monitoring — is the most easily overlooked. Many people don't realize their context has exploded until model quality drops off a cliff. This is one of the core reasons I built GroAsk — one glance at the menu bar shows every Claude Code terminal's context usage, current status, and queue state. No more switching between windows to check.

Multi-Session Management: Stop Staring at One Window

Don't run just one Claude Code process at a time — watching it work is too inefficient.

  • Comfort zone: 2-3 Claude Code processes in parallel
  • High-intensity: Scale up to 5-10
  • Monitor each session's status in real-time, only attend to sessions that need user input

But here's the practical problem: with 5+ terminals open, just finding "which session is waiting for me" via Cmd+Tab is painful enough. My approach is using GroAsk's dispatch panel to centrally manage all Claude Code terminals — which is running, which needs input, which is almost out of context — all visible at a glance. ⌥Space to summon the panel, one click to jump to the right terminal.

If you're also a multi-session power user, give it a try: groask.com

Where Are the Model's Limits?

The model's boundary is truly novel creation.

Building something no one has built before — it's not in the training data, only inference. That's where human creativity and judgment matter most.

AI can execute efficiently, but the decisions of "what to build" and "how to build it" remain human territory.

Summary: Three Takeaways

  1. Claude Code isn't an AI assistant — it's an AI team. Manage it like you'd manage a team.
  2. The human's core value is decision-making. Product decisions, technical decisions, innovation — these can't be outsourced.
  3. Engineer your AI usage. Skills codify workflows, Subagents manage complexity, knowledge grouping enables memory, context management preserves intelligence.

The author uses Claude Code daily to build GroAsk — a native macOS AI workbench. ⌥Space to reach all your AIs, with real-time monitoring of multiple Claude Code terminal states. If you're using Claude Code, give it a try.