Claude Code is really poorly named. It’s awesome, powerful, a game changer, but it’s most powerful use cases are not for writing production code.

Instead, be a surgeon, and use Claude Code as an assistant.

The Surgical Team

In the legendary book The Mythical Man Month, Fred Brooks introduces Mills’ the surgical team. The idea is simple, one master surgeon who does the core work (production code) supported by assistants (who do literally everything else). The reason for it is because:

• Production coders all need to communicate with each other (O(n^2) runtime), but their assistants don’t, so you scale up the assistants to keep n down.

• Bad programmers add too much tech debt, but that might be all you can hire.

• If a 0.5x programmer can improve a 10x programmer by 10%, that’s on net better than him contributing directly.

This fai…

Read more

When you read the Unix guys, they talk a lot about this, that 90% today is much better than 100% tomorrow. And this is still true today. I have a draft of an essay that I was working on about coding with Claude Code, about how Claude Code is actually not great for production coding, but is an excellent assistant and therefore you can code with surgeons. And then, as I was working on this and refining it and crafting it and getting it all perfect, someone else came out with the exact…

Read more

We’ll be coding faster with claude code.

Except, not much has changed.

Fundamentals still matter. Expertise still matters.

That’s what Coding Faster is really about, but instead of doing it through programming language directly, now sometimes we do it through claude code.

Read more

Read more

In particular, there's one theme that always comes up, which is that the new technologies allow you to do things without understanding what you're doing. In the short term, this is fantastic. Now you can do things that you didn't know how to do before, and you don't need to pay t…

Read more

You aren’t God.

You are a programmer. Use a language that maps to how you think. That’s fast.

Using an unnatural language? That’s slow.

Plus of course python has way more library support, stackoverflow support, reddit support, friend support, better gpt support, and programmer market liquidity.

JaneStreet of Ocaml fame has had to write everything themselves in Ocaml…because nobody else wrote it in Ocaml. That’s slow. Yes they’re good, but not because of Ocaml. At best, you could say a Prop shop needs to write everything anyways for security reasons so its less slow, but I doubt even that’s true.

Now functional programming ideas. Like map, filter, reduce. Those are great, go study them. I highly recommend SICP, posibly the greatest book on programming ever written. It uses LISP, but only as a teaching device to express the books ideas.

Read more

So you might think, oh that standard sucks, it’s opaque. I’m going to make my devs lives better and make a human friendly version.
That’s what the devs who designed moralis did:

https://docs.moralis.com/web3-data-api/evm/reference/get-token-metadata

instead of 1 we have eth. eth, I know what that means!

instead of 137 we have polygon, I know what that means!

Except now to use their api with my data, I have to extract their custom list of chain names and map my standard list chain ids to their custom standard that nobody else supports.

And if every api did this, then I’d have to do a version of this for every api. O(n) dev time

If no api did this. I’d never do this. O(1) dev time

Chain ids might be a bad standard, but it …

Read more

if(!valid(value)) {
  console.log("Invalid value")
}

so people must be writing slop like that, let’s fix it:

if(!valid(value)) {
  console.log(`Invalid value:${value}`)
}

Isn’t that so much nicer! Now we can see in our logs what the value is. And it’s DSV so we can parse it with cut if we need the value (usually in my paste buffer but occasionally elsewhere).

Sometimes I’ll write a sick little:

./run | tail -n1 | cut -d : -f 2 | pbcopy

or if I log JSON

console.log(JSON.stringify({s: "Invalid value", value})

I can get the value via:

./run | tail -n1 | jq .value -r | pbcopy

isn’t fast programming fun?

Read more

Yes..and..no..and..no.

The intended difference is that taped together works under demo load but not production load where as “built” does. But there’s a secret to taping stuff together that does handle production load.

“Hacky Js” Should work in theory. In 2024, computers are fast. You shouldn’t need “industrial strength code” or to spend time optimizing performance to get something that’ll scale. It should be possible to throw something hacky on a scaling system like Vercel or aws and It Just Scales. But when you do…something breaks. Something always goes wrong. So how do you pull off the fine art of shipping hacky js?

Get good.

One really great thing about getting good at fast programming is that you realize being good at fast programming is mostly about not making fatal mistakes. Don’t violate SPOT. Keep the code and data trans…

Read more

Why code for the future?

Taking a big step back, when we code for the future, it’s because we want to make sure our future selves don’t have more work because of bad decisions today. We’ve all accidentally made technical debt that made extra work for ourselves. But many times, we’re so worried about the future, we overcorrect. We make the mistake of trying to predict the future so correctly that we code today so that tomorrow is less work because of good decisions today. That’s way more than we need to strive for, and it’s extremely difficult. Instead, let’s just shoot for “not more work” by not making tech debt.

How to not code for the future

This means, we should write code that doesn’t box ourselves in and STO…

Read more

I just copy pasted the above, with this simple script:

#!/usr/bin/env bash
scrot -s -e 'mv $f ~/Pictures && xclip -selection clipboard -t image/png -i ~/Pictures/$f'

Ask AI for help installing any dependencies.

If you have i3, you can bind it, I bind mine to mod+Shift +p and save it as screen-captue

bindsym --release $mod+Shift+p exec screen-capture

Then, try it out.

1. Execute

2. Click

3. Drag

4. Click

5. Paste

The Image will be dropped both into your clipboard and ~/Pictures

That’s it.

Read more

1 - Go find one that’s been already made

2 - Make a new one

We made a new one because it’s much much faster. Finding one requires searching codebases, extraction, and maybe, even, dealing with people. It also could requiring verifying that there’s no pollution issues with reusing the old one.

All of which, especially people are slow. But creation? That’s fast.

So why would anybody even suggest reusing the first one?

Scarcity mindset.

The scarcity mindset is toxic for fast coding. If you suspect you have it, you must root it out, remove it, and develop the Abundance Mindset. Stop thinking everything is expensive and you must think hard about how to conserve it. Instead everything is cheap and you must instead think hard about how to lay it out cleanly so you can work quickly us…

Read more

The Options

Today, there’s still only the big 3: Windows, Mac, and Linux. non-mac BSD is esoteric. Temple OS doesn’t support networking. And Plan 9 never made it out of the lab. The big 3 are all popular enough that we won’t go into them or their differences.

The Golden Rule

The golden rule for choosing any long term personal technology choice like an OS or a text editor. Once you’ve ruled out everything that doesn’t meet the technical bar, we should pick:

The one that you’ll want to spend the mos…

Read more

1. prompt better like you would to a human.

2. Tell the LLM to be good at it’s job (fun quirk of LLMs is that it does the average of what it’s seen, so careful prompting can deal with slop)

and I don’t need to tell you that. Most of my usage has been the obvious usage, and since I had nothing novel so say, I said nothing. However, lately I’ve started to come across some non-obvious ways to use LLMs.

Surgeon style programming

We can reorganize ourselves as a head surgeon from the surgical team in The Mythical Man Month. The book, from 50 years ago discusses that every surgical team needs a human language lawyer, human toolsmith, etc. but getting people to agree to that was hard…

Read more

Just kidding, I’m moving onto neovim, the latest and greatest vi derivative, here’s why:

vi/vim have served me very well for text editing and integrate nicely into the unix shell. However…vi/vim are not made to be extensible. vimscript is an awful hack. They’re made to edit text and they do that well, but nothing more, which in the 80s and 90s was the right choice.

But now, in the 2020s, text editors need some level of extensibility to handle integrations with things like static analysis tooling and genAI. Text is not just edited in isolation, but in coordination with feedback from various daemons. Here, neovim shines as neovim is an extensible vim. A vim you can plug things into.

By 2024 i…

Read more

CToD(BToC(AToB(A)))

Good code:

DFromC(CFromB(BFromA(A)))

Read more

NOPE.

Here’s the query that solved my problem faster:

make this nice json

followed by the raw data.

Just try the AI first, and don’t be clever.

If you have too much data for AI, then sure, a parser’s great. but until then, try the AI.

Read more

Install this custom script with TamperMonkey.

// ==UserScript==
// @name         Dark HN
// @match        https://news.ycombinator.com/*
// @run-at       document-start
// @grant        none
// ==/UserScript==

(function() {
    'use strict';

    // Create a style element
    const style = document.createElement('style');
    style.innerHTML = `
        body {
            background: #111 !important;
            filter: saturate(0) invert(1) hue-rotate(180deg) !important;
        }
        td {
            background: #ccc !important;
        }
        img {
            filter: saturate(0) !important;
        }
    `;

    // Append the style element to the head
    const head = document.head || document.getElementsByTagName('head')[0];
    if (head) {
        head.appendChild(style);
    } else {
        // If the head element is not available yet, use a MutationObserver to wait for it…

Read more

Nope.

This works in a narrow use case. You write a change and commit everything you just wrote. This is, in my experience is not at all how I use git. I’ll make a bunch of changes and commit…some of them. There’s 2 reasons I do this:

• I’ll never commit some changes like adding logs, custom builds, removing flaky tests

• I break up my changes into lots of small commits. Often times my creation process is super messy. I don’t really know what I’m doing until it’s done. But once it’s done, then it’s clear how to break the change up into bite size commits to make a clean history that’s easier to bisect and analyze for future developers.

Prettier sadly works on not the diff bu…

Read more

1. PSQL can handle all your data

2. PSQL can handle all your data operations

A simple difference

The difference between data and operations can be explained with two examples:

example 1 I need to store customer records and have a message queue for incoming requestions. I could do it in psql for storage and redis for queuing, or just psql.

example 2 I need to store customer records and run a cron job on them daily. I could do it in psql for storage and use a system cron, or I could do cron within psql.

PSQL can handle all your data

The claim is you don’t need to use other dbs, you can do it all with psql.

This is awesome for a fast coder. You don’t need to use a different technology for things like nosql, graphql or message queuing. We can use one technol…

Read more