The honest take
Claude Code has moved from "an interesting CLI" to "the most reliable AI pair programmer we use" inside Kiebot. The improvements are not flashy. They are about staying out of your way.
What we actually use day to day
- Long-context coding. The 1M context Opus tier means we can load the full module under review without trimming. It changes how we triage bugs.
- Subagents. Background agents handle research, repository search, and routine refactors while we keep typing in the main thread.
- Skills. Custom skills for our internal flows live in the repo. New engineers get the same shortcuts as veterans on their first day.
- Hooks. Pre-commit and post-task hooks let us enforce conventions automatically.
- Memory. Persistent project memory means we no longer re-explain our codebase every session.
What we keep switched off
- Auto-commits. We still want the human in the loop for the commit message.
- Auto-push. Pushing remote is a deliberate act.
- Anything that touches production credentials. No exceptions.
The unlock most teams miss
Skills and hooks. A team that writes three or four shared skills for its codebase will outpace a team that uses Claude Code with default settings. The tool becomes specific to your project.
Where we still review carefully
Database migrations, authentication flows, anything money-related. The model is fast and confident; humans still have better judgement for blast-radius decisions.
Related reading
For the comparison side, see Gemini CLI vs Claude Code. For the broader AI-assisted coding mindset, see The AI Pair Programmer.
