Claude Code shipped a feature called /insights. It generates a personalized report of how you actually use it - what you build, how you build it, where you get stuck, and patterns you don't notice about yourself.
If you want to get the most out of the tool, this is it. Not tutorials, not docs - a mirror of how YOU work.
I ran mine after 43 days of daily use. Some of the numbers were expected. Others surprised me.
The Numbers
4,507 messages across 149 sessions. 53,664 lines of code written. 420 files touched. Median response time: 53 seconds.
But the number that caught me: 1,653 of those messages happened between 6pm and midnight. An average of 104 messages per day, mostly after hours. Night owl who lives in the terminal, confirmed.
How AI Sees Me Build
The report described me as a "vision-driven iterative refiner who provides minimal upfront specs and relies on rapid correction cycles."
It broke my sessions into types: 27 were iterative refinement (building one thing through rounds of feedback), 16 were multi-task sessions (jumping between projects), and the rest were quick fixes or exploration.
My top request categories: image generation and community support drafts. Not what I expected. I thought it would be database queries or UI work, but apparently I spend the most energy on visuals and helping people.
It also said I treat AI "like a co-founder who handles execution across every layer of the stack" - from database to design to deployment.
Honestly? I didn't think about my workflow that way until I read it. The "minimal upfront specs" part hit different. I know Claude Code has deep context on my projects, so sometimes I skip the brief entirely and expect it to just know what I mean. That's on me. And now I know it's a pattern, not just a one-time thing.
The Supabase Thing
This is the stat that surprised me most: 258 Supabase SQL queries in 43 days. All from the terminal through MCP. Zero from the Supabase dashboard.
I was sharing this with someone from the Supabase team this week - how I don't really visit the dashboard anymore. RLS policy checks, schema changes, data integrity audits, blog post management, even creating new projects. All from the terminal through Claude Code.
The MCP integration changed how I work with my database completely. I describe what I need, Claude Code writes the query, executes it, and shows me results. No context switching. No browser tabs. I basically live in the terminal now.
Multi-Clauding
The report tracks something called "multi-clauding" - when you run parallel Claude Code sessions at the same time.
I had 82 overlap events. 24% of my messages happened during parallel sessions. One session building UI components, another querying Supabase for schema changes, another drafting community support responses for Discord.
I thought I was just switching tabs. Turns out I was running a parallel workflow without realizing it.
Learning How to Use Agents From How I Already Work
This is the part I didn't expect.
The report didn't just show me stats. It showed me where my workflow could level up. Based on the tasks I do most - querying Supabase, generating images, managing content across projects - it explained how I could use parallel agents to handle multiple things at once.
And once I understood the pattern, I tried it. I ran 4 sub-agents simultaneously to audit my database schema. One checking table structure, another analyzing RLS policies, another reviewing query patterns in my code, another scanning for orphaned data. All running at the same time, each reporting back independently.
That's something I wouldn't have tried if the report hadn't shown me that I was already working in parallel without knowing it. It connected the dots between what I do and what's possible.
The Flaws
The report doesn't just tell you how you build. It shows your friction.
My top issues: 60 misunderstood requests and 57 wrong approaches. But here's the thing I didn't expect - I correct tone more than code. Most of my back-and-forth isn't about fixing bugs or logic. It's about getting the voice right. "Too formal." "Less corporate." "More like how I actually talk."
Part of it is speed. I use Whispr for voice-to-text and sometimes I'm going so fast that misspellings slip through - and then I'm spending rounds fixing something that wasn't a real instruction, just a typo the AI tried to interpret.
My image generation sessions average 6-12 rounds each. The report's "fun ending" section literally called me out for spending 12+ rounds trying to get the perfect Instagram story aesthetic before giving up and starting a new session.
Fair.
What I Did About It
The report gives you specific, actionable fixes based on your patterns. So in one session, I implemented everything it suggested:
- Custom skills with my exact tone rules and brand aesthetics
- A TypeScript type-checking hook that runs automatically when a session ends
- Workflow templates for repetitive tasks I was doing manually
- A full Supabase schema audit with 4 parallel agents - found 6 critical issues in 5 minutes, including an INSERT policy that was named correctly but implemented without any actual role restriction
The Unexpected Part
Using Claude Code daily has made me understand technical concepts I didn't get before. RLS policies, database indexes, TypeScript type checking, edge functions, JWT flows.
Not because I sat down to study them. Because I use them every day through conversation. The repetition builds understanding naturally. The terminal went from being intimidating to being the place where I concentrate best.
The Reverse
And if you think about it, that's kind of the reverse of how this usually works. We spend all day correcting the AI - fixing its tone, steering its output, telling it when it's wrong. But here, the AI is the one giving you feedback. Showing you where you're inefficient, where you repeat yourself, where your process breaks down.
The AI coaching its human. I didn't expect that to be useful, but it was.
Three things I'd suggest after running yours:
- Save it. The report is a snapshot of how you work right now. It's worth keeping.
- Make it a habit. Run it periodically. Your patterns change as you get better, and the report will reflect that.
- Actually act on it. The suggestions are specific to your data. I implemented mine in one session and my workflow changed immediately.
43 days in, and honestly the terminal is where I concentrate best now. Didn't see that coming.