Claude Code $200 Plan: Lessons from 30 Days of Heavy Use

Executive overview

A developer spent $200 on Claude Code's Max Pro plan for 30 days and documented where the money was wasted and where it wasn't. The core finding is that the $200 plan is overkill for anyone coding less than four to five hours daily — the $20 plan covers most users. Beyond the pricing question, several workflow decisions significantly affect output quality: model choice, context management, checkpointing, and preparation all matter more than raw token limits.

Sonnet 4 outperforms Opus 4 for coding, and context hygiene matters more than model tier.

Plan selection

• $200 Max Pro plan is only justified at four to five hours of active daily coding
• Most users are well-served by the $20 plan; upgrade to $100 only if limits are hit
• Start at $20, observe usage, escalate only when necessary

Model choice

• Opus 4 is the largest model but does not consistently produce the best code
• Sonnet 4 outperforms Opus 4 on instruction-following, one-shot prompts, and aesthetics
• Switch model via /model command in the terminal — set it to Sonnet 4 explicitly

Context window management

• AI intelligence degrades as context window grows — keep it short for best output
• Two built-in options: /clear (full wipe) and /compact (compressed summary)
• Prefer /clear over /compact — compact's output is opaque and unverifiable
• If context is already bloated, ask the AI to summarize only the specific things needed (e.g., recurring errors, failed fixes), copy that output, run /clear, then paste it into the fresh session
• This gives full control over what carries forward

Checkpointing with Git

• Claude Code has no native checkpoint/restore feature unlike Cursor or Windsurf
• Manual Git workflow is the substitute: git add . → git commit -m "message" → git push
• Run this after every validated micro-feature: build → test → manually validate → commit
• When AI breaks the codebase, point it to the remote repo URL and have it reset to the last good commit

Sub-agents

• Sub-agents are useful but should be used sparingly — parallel agents cause conflicts in interconnected codebases
• Cognition (Devin) recommends single-threaded agents for coding; parallel agents found to be slower in practice
• Good use cases: a specialized UI agent with a focused system prompt, or a dedicated debugging agent with a fresh 200k token context window
• Fresh context window is the key benefit — the sub-agent is not polluted by the main thread's history
• Avoid spawning multiple sub-agents simultaneously until verification layers exist

Using Cursor alongside Claude Code

• Keep a $20 Cursor subscription even when using Claude Code — it provides access to o3 and Gemini 2.5 Pro
• Claude repeatedly failing to fix a bug can often be resolved immediately by passing the same problem to o3 in Cursor
• Different models have different strengths; no single model wins on every problem type
• Cursor also provides superior linting and codebase mapping features

CLAUDE.md rules file

• CLAUDE.md acts as a persistent rules/context file the AI reads before every action
• Avoid /init — it crawls the whole codebase and generates an over-bloated, quickly stale file
• Use a retroactive approach: add entries only after solving a recurring error, not upfront
• Ask the AI to summarize the error, why it kept failing, and what future guidance to add — then append that to CLAUDE.md
• Practical examples: storing Supabase credentials with access instructions; saving Playwright MCP login credentials for end-to-end tests

Preparation process

• 70% of project time should be preparation; 30% execution
• Three documents to create before coding:
  1. Specification document — the "what," generated via AI reverse-interview (AI asks one question at a time, extracts requirements)
  2. Blueprint — the "how," converted from the spec
  3. To-do list / roadmap — converted from the blueprint; AI checks off tasks as it builds
• More preparation effort directly reduces coding errors and iteration loops