My 10x iOS development workflow with Claude Code

Opus 4.5 has been a step function improvement on what was previously possible with coding agents to where it is not uncommon for developers to write 90%+ of their code using AI. This is only possible with proper context management, good prompts, and an up-to-date CLAUDE.md file for your project.

My Claude Code workflow

1. Always start with plan mode (shift + tab) to describe a feature and @ reference all relevant files in my codebase. I typically dictate with Wispr Flow since I can speak at 165 words per minute and only type at 85-100. I’ll attach a screenshot of similar UI from Figma or a competitor if the front-end is complex.
2. Answer questions about the implementation with Claude.
3. Claude starts on the implementation. While Claude is working, I’ll sometimes split my terminal and open another Claude Code instance in the codebase and repeat steps 1-2 for a feature that doesn’t conflict with the current implementation.
4. When finished writing code, Claude will typically do an Xcode build to make sure everything compiles. I’ll then build the changes and review them on my phone. If something is wrong, I’ll send follow-up messages to Claude to fix it.
5. Once everything is working properly, I’ll tell Claude to run the code-simplifier agent which will spin off a subagent with a fresh context window to review all of my recent changes for inefficiencies.
6. If any new strings were added during the changes, I’ll run my /localize command to look at the string file diff and add all necessary translations for my supported languages.
7. After code-simplifier has cleaned up the code and translations have been added, I’ll run my /commit or /pr commands depending on if I’m ready to deploy. These commands will automatically write commit messages and open a PR using the GitHub command line interface.
8. If comments are left on the PR in-line by Cursor bug-bot or a human, I simply prompt Claude to address the comments left on the PR and it will pull in the information automatically using the GitHub CLI.
9. Depending on how large the change is, this should likely be done in a new context window. Check by running /context. If context > 35% then run /compact to make the context smaller and address the comments or run /clear to start with fresh context. Claude gets noticeably dumber if the context is > 60% filled.
10. After implementing this feature, I then ask myself if there was any manual work I had to do or prompting that could have been resolved if Claude had additional context in CLAUDE.md. If yes, then I’ll ask Claude to update it from my past session. I use specific instructions to keep CLAUDE.md clean and concise.