Claude Code is a chainsaw | Mechanical Survival

Claude Code is a chainsaw

13 Jun, 2026

“Wood, chainsaw, tree” by pb826 on Pixabay

I first picked up a chainsaw two summers ago.

I was staying with a friend at his farm on Dartmoor. We spent a drizzly<br>afternoon splitting logs with a hatchet, clunking and thunking our way through<br>the pile until the time came to carve up a bigger section. He said “I’ll tell<br>you what”, grinned, disappeared into an outbuilding and came back brandishing a<br>green and white chainsaw with a long black blade. Battery-powered, unfortunately.

To use it, he showed me, you first flip up the guard then gun<br>it in the log’s direction, cutting the air before the wood. Once it bites, the<br>course is set and your arms must keep it moving.

More than once I got stuck because I underestimated the force required. I had<br>expected the work to become suddenly trivial, but the system of cutting, for<br>all the chainsaw’s speed, was not fundamentally different from how it had been<br>with the hatchet. It was still knotty fibres and fast-moving steel. Except now<br>with chewing rather than slicing, some extra danger, and different muscles<br>required.

I recognised this feeling when I recently took up Claude Code Max to do<br>business-critical work with it.

Cutting

Claude Code Max is a chainsaw.

I&rsquo;ve used Claude Code&rsquo;s Pro tier before, so I have an idea of what I&rsquo;m doing. I<br>took time to flip the guard up, so to speak: I gave it a grounding in my<br>preferred architectural patterns and a strict testing regime. I then got it to<br>proceed on my domain, one user story at a time.

The real chainsaw did better the faster you could get it to cut. And as I made<br>my first fast, confident cuts with Claude I marvelled at how it ate tirelessly<br>and tastefully through every task I threw at it. But in Claude’s case too, I<br>found new muscles were required.

One way of going wrong is like getting the saw stuck. Claude Code eats and eats<br>until it finds a knot (really, some problem that it can’t deal with) and the<br>knot just will not yield. But the blade still mindlessly turns, blunting itself<br>as the context window (and the codebase) fills with rubbish.

What happens after I stop cutting is also new, but more interesting. Here the<br>risk is not so much stopping progress as the whole piece of work falling<br>apart. In the wake of Claude’s chainsaw there’s sloppy microcopy, UI jank, and<br>various kinds of subtle incoherence—a security vulnerability here, an oddity in<br>the data model there—that can undermine the whole product.

Unfinished work is now my enemy, and it accumulates very quickly under the<br>blade. My instinct is to take very deliberate cuts, usually user-story-sized,<br>and force myself to put down the chainsaw to finish each one properly.

Finishing

Finishing is not “reviewing the AI’s work”. That’s a cliché that treats AI like<br>a person, and the activity as no more than a stage-gate.

But finishing isn’t checking. It’s wielding the sander; it’s its own<br>discipline. It’s realising that the user management screen you’ve just sawed<br>contains a bizarre assumption about what non-admins will see and then having to<br>redesign it there and then because all the new curves in your product were<br>made by a tool that just wanted to go in a straight line.

When finishing, I rarely begin with code. Revisiting Claude’s first rugged<br>cuts, I actually use the software itself. This is strange to me—in the past I’d<br>know it inside out by the time I completed a piece of work. I have to learn it<br>anew through each change. I explore the design, code structure, and every other<br>small aspect of the system that falls under the scope of the story I’m working<br>on.

Then I return again and again to the tool to make ever-smaller changes. When<br>it’s smooth, I start again.

This takes time, but if I skip these finishing phases I feel like I’m losing my<br>grip: scope continually expands just beyond reach, and the fear of something<br>going wrong never leaves me.

Life and limb

When the chainsaw gets out of control, kinetic and spectacular injuries ensue.

The faster and further I cut the faster I must decide and the more decisions I<br>make. I can easily rack up more than I can finish before I must cut again to<br>keep moving. That cost compounds in technical debt: less splatter, but no less<br>lethal.

Holding the nerve to keep stopping and stopping when the tool just wants to eat<br>is hard. And I don&rsquo;t like that much of the time it&rsquo;s just me and the<br>single-minded agent. Working one-to-one with it feels like a lot of<br>responsibility for me as individual, and somewhat disrespectful of its<br>seductive power.

So I find myself wanting other people physically beside me at the console to<br>bring alternative perspectives and keep us honest. I crave, I think,<br>multidisciplinary mob programming? With only one chainsaw in operation<br>at a time.

#Coding-Agents

#Ai

#Programming

#Management