I spent $697 on Claude Code last month. Here's what I learned.
I spent $697 on Claude Code last month. Not on a course about it. Not on a wrapper tool. On raw API usage. Tokens in, tokens out. Building things. Some of that money was well spent. Some of it was burnt on stupid mistakes I could have avoided. Here's what I learned. 1. Start with a PRD, not a prompt. My most expensive sessions were the ones where I opened Claude Code and typed "build me a dashboard." No spec. No requirements. No constraints. The AI happily built something. Then I realised it wasn't what I wanted. So it rebuilt. And rebuilt again. Each rebuild burning through tokens on code that got deleted. Now I write a PRD first. One page. What it does, who it's for, what's out of scope. I paste it into the CLAUDE.md file so every session starts with context. The PRD costs me 20 minutes. It saves me hours of token-burning tangents. 2. Plan with an affordable model. Build in the CLI. Claude Code defaults to the most powerful (and most expensive) model for everything. Including planning. I stopped doing that. Now I plan in Claude.com using Sonnet. It's fast and cheap. I hash out the architecture, the file structure, the implementation steps. I ask it to poke holes in my approach. Then I take that plan into Claude Code (running Opus in the CLI) and execute it. The expensive model does what it's good at: writing code, handling complexity, managing multi-file changes. Planning in Sonnet costs pennies. Building in Opus costs dollars. Don't use dollars for planning. 3. Use multiple agents, not one mega-session. My biggest waste was running single sessions for hours. One agent trying to do everything: plan the feature, write the code, debug the tests, update the docs. By hour three, the agent was hallucinating. Referencing functions that didn't exist. Suggesting fixes for problems it created two hours earlier. Now I split the work: Agent 1: Write the feature Agent 2: Write the tests Agent 3: Review and debug Agent 4: Update docs Each agent gets a clean context window. Each one does one job well. I run them in parallel across multiple terminal windows. More agents = more tokens. But the output quality is so much higher that I actually spend less overall. Fewer rewrites. Fewer debugging spirals. Fewer "let's start over" moments. 4. /init every project. Every time. The /init command creates a CLAUDE.md file for your project. It's the AI's instruction manual. What the project does. What conventions to follow. What mistakes to avoid. Without it, every new session starts from zero. The AI asks questions you've already answered. Makes assumptions you've already corrected. With it, every session starts informed. The AI reads the file, understands the project, and gets to work. 20 minutes setting up CLAUDE.md saves hours across every future session. The real lesson: $697 sounds like a lot. But I shipped a trading CLI, a newsletter scheduling system, an AI email analyzer, and three automation scripts. In one month. As a solo developer. The cost wasn't the problem. The waste was. And most of the waste came from skipping the boring stuff: planning, context, constraints. AI coding tools reward structure over speed. Every time. Want to learn the exact prompts and frameworks I use to build faster with AI? Check out my Prompt Writing Studio course |