There is a command in Claude Code that most people do not know exists. Type /insights, and it reads your entire history of work across your projects, then writes you a short report on how you actually use AI.
Not how you think you use it. How you actually use it. The patterns, the habits, the things you keep doing that you have never noticed. Then it suggests concrete changes to your CLAUDE.md, the file that personalizes how Claude Code behaves for you.
It is, quietly, one of the most useful features for anyone trying to get past asking and into building. Here is what it does and why it matters.
The problem it solves
Everything in this series so far has shared one assumption: that you can see your own habits clearly enough to improve them.
Write your standards down. Build a judgment system. Iterate every three weeks. All of that works, but it depends on you correctly noticing what you do. And people are bad at noticing what they do. You have habits in your prompts you cannot see. You repeat instructions you could have set once. You ask for things a configuration file could have supplied automatically. You carry workarounds for problems you already fixed.
The /insights command exists because self-review has a blind spot the size of your own routine. It turns the mirror around. Instead of you guessing at your patterns, the system reads thousands of your actual interactions and tells you what they are.
What the report shows
Run /insights and you get an analysis page built from your real history. The specifics vary by what you actually do, but the shape is consistent.
It surfaces the instructions you repeat across prompts, the ones that should live in your CLAUDE.md so you never type them again. It flags patterns in how you ask for reviews or changes, which often reveal a judgment system trying to emerge that you have been running by hand. It points out where you are doing manually what a persistent instruction would handle automatically.
The output is not abstract advice. It is specific to you, because it is built from you. Two developers run /insights and get two different reports, because they have two different sets of habits. That is what makes it useful. Generic prompting advice is everywhere. A report on your actual habits is rare.
Why this turns iteration into a system
Earlier I described the iterate cadence: fifteen minutes every three weeks to keep your judgment systems sharp. That cadence depends on you doing the noticing. You review fresh work, find the misses, improve one standard.
/insights automates part of that loop. Instead of you manually hunting for the habits that are costing you time, the system generates that review for you, on demand, from data you already produced. You still make the decisions. But the surface area of what you are deciding about gets handed to you, organized, instead of you having to dig it up.
This is the asker-to-builder shift applied to your own workflow. The asker reviews their habits manually, when they remember, incompletely. The builder runs a command and gets a system-generated review of their own patterns, then acts on it.
CLAUDE.md as your standards, encoded
The reason /insights points its suggestions at CLAUDE.md is that CLAUDE.md is where your standards live for code work. It is the persistent context file that shapes every session. The commands you want followed on every task, the style you require, the review bar you enforce, all of it belongs there, once, instead of being retyped into every prompt.
If a judgment system is your standards encoded into a prompt, CLAUDE.md is your standards encoded into the tool itself. /insights is the diagnostic that tells you which standards you have been carrying in your head instead of writing down where they belong.
Who this is for
/insights is a Claude Code feature, so it applies most directly to people doing development or technical work in the terminal. But the principle generalizes. Whatever tool you use, the highest-leverage move is the same: get a view of how you actually work, then encode the repetitive parts into a persistent layer so you stop doing them by hand.
The people pulling ahead right now are not the ones with the most clever prompts. They are the ones who have stopped carrying their standards in their head, who let a system show them their own patterns, and who turn those patterns into configuration once instead of typing them forever.
Run the equivalent of /insights on your own work. Find the habits you cannot see. Move them out of your head and into the system. That is the last mile of the shift from asker to builder.