← Claude Code Hub
Jul 8, 2026 · 4 min read ✦ Tip #128

Concentric circles: the method for keeping up when Claude touches half the repo

A big or unfamiliar change is hard to grasp reading it in silence. A method to get to the bottom: dictate your review to Claude and narrow the circle, pass by pass.

Concentric circles: from the outer layer with every file to the center, shared understanding between human and model

TL;DR When a change is too big, too new, or too unfamiliar to grasp at a glance, I don't read it in silence. I start dictation and read it out loud to Claude, file by file, flagging what I don't understand. Then I iterate in concentric circles: each pass halves the area, until the model and I understand the same thing.

With frontier models touching a pile of files at once, the bottleneck has moved. It's no longer writing the code, it's understanding the code that's already written. And reading in silence doesn't work: you skim, you skip, you lose the thread. Reading out loud forces you to actually look. And if you dictate that reading into Claude Code, the model learns exactly where you got lost.

It's not a feature. It's a method. The tool is dictation: I use VoiceInk, but Claude Code ships a native one (/voice, hold Space) if you'd rather not add anything external.

Result (the passes of a real review):

Pass 1 · the whole change (12 files)
> [voice] "auth.ts, I see the guard... Claude, why do you validate
   the token in the middleware and again in the handler?"

Pass 2 · only what didn't close (5 files)
> [voice] "the refresh flow. I don't get this, where does the TTL come from?"

Pass 3 · the core (1 file)
> understanding ✓  →  next action

The method

1. Outer layer: narrate the read-through

Start dictation and go file by file saying what you see. Don't summarize it in your head: say it out loud. Most of them, or the important ones, or the essential part. The act of verbalizing already makes you look where you used to skim.

2. Flag what you don't understand and address the model

Call out every doubt aloud: "Claude, I don't understand this part," or "ok, I get why you did this, but this, why?". By the end of the pass, the model already knows your blind spots and what you think needs adjusting or refactoring.

3. Close the circle: narrow and repeat

From all of the above you move to a smaller sub-context. The exploration area drops by around 50%. You read again, you ask again, the model validates. Another layer. And another. It helps to drop the blind spots into a task list so none slip between passes, as long as your model still saves it.

4. The center

You keep going to the core: shared understanding between human and model, or the next clear action. You start on the outside, there are layers to cross and validate; the target is the center.

Reference

Layer What you do What you get
Outer Narrate every (or the essential) file out loud The model knows your doubts and suspicions
Middle Narrow the area ~50% and re-read only what didn't close The already-understood parts drop away
Center Iterate to understanding or the next action Human and model aligned on the same code

The human: in the loop, or above it?

Even though I aim for the most atomic changes I can, with frontier models like Fable you inevitably end up with a pile of files to validate. Pair this with plan mode so the change is born already sliced up. And this method now has a visual variant: instead of reading the code, Claude sculpts an artifact of the change and you carve it version by version until you understand it. I cover it in how I use artifacts to understand what I delegate. And there's a question I'll leave for another tip: should the human be at this point in the loop, reviewing file by file, or a level above? For now, when I don't understand something, I drop into the loop. And I drop in talking.

Official docs: Voice dictation

Found this useful? See more of wmedia.es in your Google results.

Add wmedia.es as a preferred source on Google
FREE GUIDE

You don't need Opus every time.

Model and effort are two settings, not one. 8 pages with the settings that actually move your bill, a three-question diagnostic for when something fails, and the price ladder.

  • 1. What "better" means capability, time and money
  • 2. Model who you send
  • 3. Effort how much rope you give
  • 4. Diagnostic three questions when something fails
  • 5. The bill where it comes from
Are you a professional Web developer?

You'll receive the guide by email · You join the Gravitas newsletter · Unsubscribe anytime

of 8
#

Wmedia · 8 pages
FREE GUIDE · 8 pages · 2 dials

You don't need Opus every time.

Model and effort are two settings, not one. 8 pages with the settings that actually move your bill, a three-question diagnostic for when something fails, and the price ladder.

Are you a professional Web developer? · Unsubscribe anytime
How I use AI

My Claude Code workflow.

The skills, agents and instructions I use as a professional Software Engineer. It's called Craft, and it's a plugin you can install.

See the skills

Using Opus for everything? See when you don't need it.