I Run Multiple Claudes at Once (and Other Surprises From My Data)
81 parallel sessions. A 72-second thinking loop. 1,140 messages after dark. The numbers behind 924 hours.
The insights report tells you what works and what breaks. The raw numbers tell you something weirder: how you actually behave.
I Multi-Claude
81 overlap events. 107 sessions involved. 13% of all messages sent while another session was already running.
I didn't plan this. Claude takes a minute to read files and run builds. I get bored. I open another terminal. Over 338 sessions, that restlessness became a workflow — feature implementation in one window, vault updates in another, code review in a third.
I treat Claude sessions like browser tabs. One isn't enough.
I Code at Night
| Time | Messages |
|---|---|
| Evening (6pm-midnight) | 1,140 |
| Morning (6am-noon) | 851 |
| Afternoon (noon-6pm) | 779 |
| Night (midnight-6am) | 14 |
The deep work happens after hours. No meetings, no Slack, just iteration. The 14 midnight messages mean I'm at least sleeping.
I Think for 72 Seconds
Median response time: 72.2 seconds. That's how long I stare at Claude's output before typing something.
The distribution reveals three modes:
- Under 10 seconds (180 responses) — "no, wrong file" / "push it" / "the build failed"
- 30 seconds to 2 minutes (572 responses) — the core loop: read output, evaluate, redirect
- 15+ minutes (127 responses) — I walked away
The productive sweet spot is that middle band. Claude produces, I evaluate in a minute, I course-correct. Rinse, repeat.
I Fix More Than I Build
| Goal | Sessions |
|---|---|
| Bug Fix | 42 |
| Git Operations | 32 |
| Feature Implementation | 28 |
| Deployment | 25 |
Bug fix is #1. Not features. Not deployment. Fixing things.
Features come in focused bursts. Bugs arrive constantly. And bug fixes are where Claude's speed advantage is sharpest — describe the problem, Claude reads the code, proposes a fix, I verify. Two minutes.
Git at #2 means I've outsourced my entire git workflow. Cherry-picks, PRs, rebases — Claude does them all.
323 Failures Per 338 Sessions
Nearly one command failure per session. Sounds terrible. It's actually the build-test-fix loop working correctly — Claude writes code, runs the build, it fails, Claude reads the error, fixes it. That "failure" is a feature.
Only 6 times across hundreds of sessions did I reject a destructive action. I trust this thing more than the error count suggests.
2 Quick Questions Out of 338 Sessions
I almost never ask Claude to explain something. Every session is a work session. "Implement the provider, hook it up, fix the bug." Not "how does this work?"
Multi-task sessions (86) beat single-task (60). I stack work. It causes more errors. I do it anyway because the throughput is worth it.
TypeScript 38:1
TypeScript: 4,139 file touches. JavaScript: 110. That's a 38:1 ratio. The JS touches are probably config files.
Markdown at 1,500 is the Obsidian vault work — a full third of my TypeScript volume, built entirely with a coding tool.
The Headline Number
81 multi-clauding events. You don't plan that. It's what happens when an AI tool stops being a thing you use and starts being a thing you assume is available. Like electricity. You don't think about turning it on — you just expect it's there when you reach for it.