What brand of AI grifter would write such a thing? With paradigm shifts comes money; and with money come the grifters. But lost among the AI grift, wealth extraction, silicon hoarding which can only be described as Smaug-like, and so forth and so forth is something beautiful, plain, earnest:

LLMs have changed my (and probably your) life, foundationally, permanently, and for the better.

All of this boils down to the fact that while LLMs are merely pretty good at writing code, they are dear-god-how-did-we-ever-work-before-this good at reading code. It turns out that having a little fella on your shoulder that can read code an order of magnitude or two faster than you can literally change your life.

Read on for joy in software, fractal toys, grifters, and at the end my tips distilled from what is in hindsight an ungodly amount of work simply not possible without this stupid little orange cube.

 ▐▛███▜▌
▝▜█████▛▘
 ▘▘   ▝▝

» software is a coastline

The core tenet of the grifter is that code’s only use is its ability to make concrete things which will produce money. Moreover, this usefulness is at worst linear. Ten times the code, ten times the money. Coding agents are often thought to produce code. The only rational conclusion is that if you’re not running six instances of Claude Code in a tmux pane¹ then – well, what are you doing?

Not so! Software, for me, is the finest toy ever created. It is better than any video game. It is better than woodworking, or LEGO, or writing music or fiction. I am a beachcomber, and software is the coastline of fractally infinite measure; I prod and scan, sure that I have finally cataloged a section, only to be surprised and delighted to find a wholly new coastline within a single grain of sand.

In other words, software has instrinsic value. A lot of it is boring, sure, but that’s the great part about the fractal nature of knowledge; the fun parts are just as infinite.

» claude is the metal detector

If I am a beachcomber, then life before June 2025 was a lot of tramping good-naturedly along the strand, mesh strainer in hand, winnowing away chaff from wheat². Life after June 2025 is like coming back from the Bobcat dealership with a brand new skid steer and purpose-built beachcombing attachment.

Thing is, we’ve had ChatGPT and friends for a few years now. The sensation of struggling to find even the terms or main techniques in some subfield are long gone. What I’m describing is different; not the broad sweeps of a cartographer’s pen, but the fine eye of a beachcomber.

Coding agents let you dig in to any piece of software extremely deeply. They are winnowing machines, finding exactly the tidbits you want. And when you get stuck understanding the tidbits, they become understanding machines, switching from searching through huge repositories for precise references to providing context on a strange looking technique, or bit twiddle, or anything else.

» why do people think claude is a dredge?

A dredge is a machine which sucks up sand from the ocean floor and spews it onto the shoreline. Usually to, like God, create something from nothing, or at least, like Moses, to make suitable for human use.

Claude is not nearly as good at dredging as it is at combing; it turns out that nearly all of the problems with coding agents disappear when using them as tools to search a specific corpus:

• They don’t hallucinate, because they have to provide files and line numbers
• They don’t destroy the codebase’s style or architecture, because they’re read only
• They can run fully autonomously, because the stakes for trying a search again are so much lower than backing up through an implementation to figure out where the wheels fell off

And when all the big problems disappear, Claude gets really, really, really good. Good enough to detect any metal worth picking up on the infinite beach we call software.

 ▐▛███▜▌
▝▜█████▛▘╲
 ▘▘   ▝▝  ╲
         (●)

» you can understand anything, now

Here’s a list of stuff that I did that I wouldn’t have been able to without Claude:

• I wrote Claude Wrapped, which lets you compare your Claude Code usage to the world, despite barely being able to stumble through setting up a static website on Cloudflare mere months prior. This thing has a god damned database!
• I gave Claude Wrapped a 3D raymarching ASCII renderer with 2 (two!) kinds of lighting, and rendered a complete winter scene featuring a Santa Claude dusted with individually lit snowflakes
• I wrote a Mermaid serializer for the internal build graphs generated by my what-if-C-had-Cargo package manager + build tool
• I did a ton of research into how good package managers resolve versions to implement my own resolver
• I wrote ~500 good, useful tests for my C standard library
• I learned how UTF8 worked, then implemented a lovely encode + decode + validate + iterate API in C
• I studied burntsushi of ripgrep fame’s lesser known glob library and implemented something similar in C
• I used the ELF spec to implement a minimal read + write + manipulation of object files
• I wrote a little code to listen for a global hotkey on Wayland, record from my Zoom, and pipe the transcription to an LLM for cleanup and then straight to my clipboard
• I jammed everything in my ~10 year old private docs repo into a vector database and made an HTMX frontend for searching through, quite literally, the corpus of my life
• I downloaded arcane blobs of JSON from Anthropic/Google/OpenAI and figured out how to reconstruct the conversations so as to embed those, too
• I built a TUI which uses the opencode SDK + OpenTUI to make a gorgeous coding agent with a context compiler in between the user prompt and the inference
• I made shitty animations of a Claude with Unicode characters for my website

I could keep going. There’s no hubris here. I am an extremely regular guy and programmer. Most of these things were handwritten code with Claude acting as my metal detector, but I just let the little fella rip, too, quite often.

Through all of these things, a few key takeaways.

»» the 0th commandment

Thou shalt run thine coding agents with auto-approve in all but the most obscene circumstances³.

»» using-llms-for-hard-stuff algorithm

1. Consider your problem deeply, and design software as you normally would. This may involve talking to an LLM about the domain, but not using it to design.
2. Find existing software which implements either a piece or most of your problem
3. Clone their repositories and ask Claude to explain deeply, concisely, and technically how they solve the problem
4. Consider their solution deeply. This may involve talking with Claude, but you must think deeply.

»» pure vibe coding works for throwaway stuff

Yeah. It doesn’t work for code that matters. To be clear, I mean the set it and forget it, Jesus take the wheel kind. No amount of eye-squinching and believing will make it so.

But for everything else? Hell yes. All the way yes. If you need some small utility, something that can be served by a couple HTMX templates, something that does a menial processing step – in other words, maybe half of the software that would be useful to you, then Claude absolutely shreds.

»» understanding libraries

Just let it rip. Any decent library has a monorepo of sorts with documentation inside the source repo. Examples, too. Documentation on the web should be dead; Claude is so far ahead of manual search that it’s not a comparison.

»» industrial fan for mental fog

Have you ever stared at some set of boolean conditions until you weren’t sure what was real? Claude hasn’t. Claude has never looked at anything. Claude would love to figure out exactly how you are a moron and cheerfully help you resolve it.

» bye for now!

That’s it. Thanks for reading. If you made it this far, enjoy this wonderful interpretation of Claude waving goodbye, as written by Claude. I have no idea what this abomination means as far as everything I just gushed about coding agents. Maybe I’m just a moron like I always was…

 ▐▛███▜▌
▝▜█████▛▘
 ▘▘   ▝▝