Claude Code Field Notes
Practical tips, setup guides, and things I've learned from daily use.
By Marco Mark. Not affiliated with Anthropic.
gh Can Attach Images Now
gh got an --attach flag in September, so agents can put screenshots in PRs without a custom skill. I found out after building one.
Sabotage-Check Your Tests
Three verification asks worth making by default: break the code and watch the test fail, run the control case, and count instead of eyeballing.
Give the Agent a Script, Not a Number
Facts written into CLAUDE.md go stale without anyone noticing. Give the agent a small script that finds out instead. Dynamic context never dates.
No Personal Skills: Keep Them in the Repo
Nobody should have personal skills. Where skills, PR process, setup and pre-PR review belong so the whole team gets them.
Using the Repo's .nvmrc in Cloud Environments
Cloud environments run on a pre-set Node version. A setup script that makes them follow the project's .nvmrc instead.
Knip: Stop Agents Writing Code You Don't Need
Agents love writing code that nothing calls yet. Knip finds it. Wire it into npm scripts, CLAUDE.md, a pre-push hook and CI so the agent cleans up after itself.
Worktrees: Isolated Agents, Isolated Work
How to use git worktrees with Claude Code to isolate agent work, plus fixes for the two biggest gotchas: gitignored files and missing dependencies.
Enabling Thinking on Bedrock
How to get thinking steps back in Claude Code when using AWS Bedrock, using the new model capability environment variables.