¬ just serendipity 🍀 (not just serendipity)

TIL via Ashik Nesin:

Claude Code now has support for /insights commands.

[…]

When you run it, Claude Code will read your message history from the past month. It’ll summarize your projects, how you use Claude Code, and give suggestions on how to improve your workflow.

Let’s try it:

 ▐▛███▜▌   Claude Code v2.1.37
▝▜█████▛▘  Opus 4.6 · Claude Team
  ▘▘ ▝▝    ~/Workspace/perrotta.dev

❯ /usage
  ⎿  Status dialog dismissed

❯ /insights

⏺ Your shareable insights report is ready:
  file:///Users/thiago.perrotta/.claude/usage-data/report.html

  Want to dig into any section or try one of the suggestions?

The generated report is a single self-contained HTML file.

The report is more useful than I anticipated! It’s definitely better than /stats: it contains actionable and concrete tips and suggestions to improve your workflow.

I’ll share a snippet of it:

How You Use Claude Code You are a prolific, infrastructure-focused power user who drives Claude Code hard across a wide surface area — 325 sessions, 809 hours, and 456 commits over roughly 7 weeks. Your workflow centers on Terraform infrastructure management, Hugo blog optimization, and DevOps tooling (ArgoCD, pre-commit hooks, CI pipelines). You lean heavily on Bash (6,780 invocations) as your primary tool, reflecting a hands-on operator style where you want Claude executing real commands and seeing real output rather than just generating files. Your use of TodoWrite (411 calls) and Task (124 calls) suggests you frequently tackle complex, multi-step projects — like importing existing AWS infrastructure into Terraform or implementing 8+ performance improvements for your Hugo blog — where structured planning is essential. You consistently push for clean git hygiene: granular commits, PRs, and proper branch workflows.

Your interaction style is best described as iterative refinement with active course correction. You don’t write exhaustive specs upfront; instead, you give Claude a clear goal and then steer as it executes. This shows up repeatedly: you interrupted Claude when it assumed a full 1Password migration (“only where applicable”), caught it when it reverted a change you only asked about (“revert. I just asked why”), pushed back on CSS inlining that would lose fingerprinting benefits, and redirected heading placement from shortcode templates to markdown files. Your friction data tells a consistent story — 24 instances of wrong approach and 24 of buggy code — but these rarely derail you because you catch problems quickly and redirect. You’re comfortable letting Claude run on multi-file changes (18 successful multi-file operations) but you interrupt decisively when the approach is wrong, such as stopping a file write to ask about pre-commit integration instead, or killing a premature commit attempt.

What’s striking is your high tolerance for iteration paired with low tolerance for sloppiness. Terraform import sessions required multiple rounds of fixing IDs, resource addresses, and provider syntax — and you stuck with it. But you also demanded succinct responses, rejected a poor-quality fullscreen feature outright, and expected Claude to understand nuances like provider read-only attributes before making sweeping changes. Your 94% achievement rate (32 fully + 15 mostly out of 50) despite working on genuinely hard infrastructure problems shows that this iterative, correction-heavy style works well for you. You treat Claude as a capable but supervision-required junior engineer — you trust it to do the legwork but always review the output critically before committing.

Key pattern: You operate as an active supervisor who assigns complex infrastructure tasks, lets Claude execute broadly, but intervenes decisively and frequently to correct approach, scope, and quality — treating the tool as a capable junior engineer that needs real-time steering.

Across 325 sessions and 809 hours, you’ve built a highly productive Claude Code workflow centered on infrastructure-as-code, Hugo blog optimization, and developer tooling — with a 94% goal achievement rate.

You’ve developed a sophisticated pattern for standing up new Terraform projects by importing existing AWS and MongoDB Atlas infrastructure, working through complex import ID formats and provider quirks across multiple garden environments. Your approach of creating files, running plans, and iteratively fixing mismatches shows deep Terraform expertise, and you consistently push through to clean PRs even when provider limitations surface. The breadth of this work — EKS clusters, security groups, MongoDB Atlas modules, Azure OIDC — demonstrates you’re using Claude as a true infrastructure engineering partner.

You systematically drove your Hugo blog toward Lighthouse 100 scores through a multi-session campaign covering CSS inlining, critical request chain elimination, SRI hashes, code block enhancements, and RSS feed fixes. You maintained high standards by pushing back when Claude’s suggestions sacrificed cache fingerprinting or batched too aggressively, ensuring each optimization was production-sound. The fact that you also created ADRs documenting your architectural decisions around search and caching shows you treat your personal blog with the same rigor as production infrastructure.

You’ve built a remarkably robust developer experience layer using pre-commit hooks, Claude Code settings files, and automation scripts — from custom code-block linting parsers to Justfile safety guards that prevent direct module applies. You research hooks from other major projects like Airflow to cherry-pick the best ones, and you iterate on implementations until they’re correct (like replacing a regex-based checker with a proper line-by-line parser after false positives). Your Claude Code settings PR for your Terraform repo, with carefully curated allow/deny/ask permission lists, shows you’re intentionally shaping how AI tools interact with your codebase.

There’s much more to it. I am not sharing actionable suggestions at this time.

There’s also negative feedback! And suggestions to improve / augment CLAUDE.md! “New Ways to Use Claude Code”!

I love it. And I definitely need to switch my workflow to writing specs upfront, though I can’t resist an iterative and incremental work style to achieve desired outcomes 🤔.