I’ve been using Claude Code steadily for the last three-plus months. I wouldn’t call myself a power user, but I’ve settled into a rhythm that works: tight PR loops, lean on it for real feature work, push back when it wanders.
Why I Finally Bothered with Claude Code Insights
The tooling in this space moves so fast that I’ve stopped chasing every new feature. I batch the exploration, maybe every few weeks to months, and see what’s still talked about. On one of those cycles, I finally got around to running /insights.
I went in mildly skeptical. I have tickets to ship, and I’m not looking for another thing to tune. But the Claude Code insights report surprised me, mostly because it held up a mirror to things I already half-knew.
Insights from the Report
The big insight? In certain kinds of UI work, I kept ending up in circular loops where Claude guesses at an API, breaks something that was working, tries a different guess, and breaks something else. I’d chalked these up to bad days. Turns out it’s a pattern, and it shows up any time it’s cheaper to guess than to check the docs. I’d just been absorbing the cost one session at a time.
Second: I’ve had to ask for reverts more often than I’d like because Claude tends to take an overly-scoped first approach. Small asks turn into sprawling diffs that touch adjacent code I didn’t want touched. I catch it, but catching it costs time and breaks my review flow.
Third, smaller stuff: skipped verification steps, conventions that quietly drift, a handful of long sessions that got truncated when I should have just broken the work up.
The Small Changes I Made
None of this was news exactly, but seeing it aggregated across dozens of sessions made it actionable in a way that individual frustrations weren’t. So I made a few small additions to my project’s
CLAUDE.md:
- A scope discipline note: minimal diffs, no adjacent refactoring, no lint-suppression as a fix.
- A reminder to match existing file conventions rather than drift.
- A verification line: actually run the checks before claiming done, and report real numbers.
- A note to consult library docs before touching config for the trickier integrations.
What I Skipped
No hooks, no subagent swarm, no autonomous scaffolder. The report pitched some of that, and it’s genuinely interesting, but I wanted changes I could make once and forget about. A few bullets in a markdown file is the right dose for me right now.
Fifteen minutes to get a clearer read on my own patterns, and a handful of lines in a config file to act on them. If you’ve been using Claude Code for a while and haven’t run /insights yet, it’s worth the look.