Developers who upgrade to Claude Max - Anthropic's $100/month tier with extended context and higher usage limits - tend to hit the same wall around week two. The tool works brilliantly for greenfield tasks like scaffolding a new service from scratch, then starts producing bloated or hallucinated code the moment you point it at an existing codebase without enough context.
The pattern is predictable. A developer describes a feature in plain English, gets back a working implementation, and assumes the tool understands their entire project. Then they ask for a modification to an existing service and get back code that ignores their authentication middleware, invents database columns that do not exist, or duplicates utility functions already in the repo.
The fix is not complicated, but it requires breaking habits from smaller-context AI tools:
- Feed it the actual files. Claude Max's larger context window (roughly 200,000 tokens, or about 500 pages of code) is useless if you only describe your codebase in words. Paste the relevant source files directly into the conversation.
- Break complex changes into sequential steps. Instead of "add notifications to the user service," try "first, read this service file and explain its current architecture" followed by "now add a notification method that follows the patterns you see."
- Use projects or system prompts for persistent context. Claude's Projects feature lets you attach files and instructions that persist across conversations, so the model always knows your stack, conventions, and constraints.
- Treat it as a junior developer, not an architect. Give it specific, scoped tasks with clear boundaries rather than open-ended feature requests.
The larger context window is genuinely useful, but it is a tool that rewards deliberate input over casual conversation. Developers who structure their prompts around the actual code - rather than abstractions of it - report significantly better results.