If you invoke the /insights command, you can get useful… insights from Claude about how you use it.
Here is mine:
___________________________________
You are a prolific Delphi developer who uses Claude Code primarily as a code review and bug-fixing workhorse across a large, mature codebase. With 222 sessions and 458 hours logged in just one month, you maintain an intense, almost industrial pace — heavily favoring systematic code reviews (30 sessions) and bug fixes (24 sessions) over greenfield feature work. Your workflow is distinctive: you typically point Claude at Delphi source files, request comprehensive reviews with bug analysis, unused code removal, and comment corrections, and then let it grind through massive Edit (4,398) and Read (3,229) operations. The exceptionally high TodoWrite count (1,367) suggests you use Claude to build structured task lists and track findings methodically, treating it almost like an automated audit pipeline.
Your interaction style is hands-on and corrective — you don’t hesitate to push back when Claude gets something wrong, and the friction data tells a vivid story. You’ve corrected Claude Code multiple times when it repeatedly fixed the wrong root cause (like the button bug where you had to intervene several times), and explicitly redirected it. When Claude’s RainDrop optimization made things slower, you knew immediately to revert.
This pattern of domain-expert oversight means your ‘mostly_achieved’ outcomes (60%) aren’t failures — they’re sessions where you steered Claude past its blind spots to get usable results.
Despite 15 dissatisfied moments and 25 ‘wrong approach’ friction events, you keep coming back at massive volume — indicating you’ve found Claude genuinely valuable as a tireless code reviewer that catches real bugs even if it needs frequent course correction on Delphi-specific patterns.
What’s working:
You’ve built a disciplined and repeatable code review pipeline across your Delphi projects — having Claude systematically find real bugs like dangling pointers, always-false returns, and type mismatches is genuinely high-value work. Your heavy use of structured task lists to orchestrate complex multi-file sessions (plugin refactors, licensing upgrades, multi-bug-fix sweeps) keeps Claude focused and
productive in ways most users don’t achieve. You also know your Delphi codebase deeply enough
to push back when Claude misreads language-specific idioms, which turns sessions into a productive collaboration rather than blind trust.
Impressive Things You Did →What’s hindering you:
On Claude’s side, the
biggest issue is jumping to fix the wrong root cause — you’ve repeatedly had to endure 2-3 rounds of incorrect fixes before Claude lands on the real problem, especially with state management and settings persistence bugs. Claude also over-flags Delphi-specific patterns as issues and sometimes starts editing before finishing the analysis you asked for. On your side,
Claude would benefit from a richer CLAUDE.md with Delphi conventions and build system details — several sessions hit avoidable compilation failures or false-positive findings because Claude lacked that ecosystem context.
Where Things Go Wrong →___________________________________
So, putting it simply, Claude says: “you push me a lot, but I am not capable enough” 🙂
Don’t worry my friend… you will get there in a few years.
