Chris

engineer turned web dev
co-founder & CTO @ - building AI for commercial construction

[Chris @_khrus]

The anti-slop ai coding prompt I’ve found the most success with is definitely “deepen modules and simplify interfaces”

It does such a good job with deduplication and centralization of core functionality, and gets the ai implementing much better levels of abstraction

It’s led to tons of consistency improvements, bug fixes, and, most rewardingly, reduced LOC. I run a deepening pass on every PR, and a nightly automation to tackle overdue maintenance

Comes from @JohnOusterhout’s Philosophy of Software Design (highly recommend), which I had read but didn’t fully leverage until @mattpocockuk put me onto it

[Chris @_khrus]

Can’t tell if I like or don’t like the “projects” or “chief of staff” agent abstraction

Feels like I’m letting myself think less and am too far from the fun

[Chris @_khrus]

It’s so fun to wake up everyday and work on my own business, work with today’s coding agents, and do engineering through a new lens

[Chris @_khrus]

Tag the incumbent in your industry and tell them why your startup with a fraction of the team size is going to eat their lunch

AI has turned small, strong dev teams into absolutely cracked producers that can disrupt legacy sitting duck companies whose software can nearly be replicated with a single prompt. Their laziness and bureaucracy is our opportunity

Here’s mine:

@BldgConnected, count your days. Anyone adopting your software today has no real sense of productivity or progress. @consightAI is going to win your market share

[Chris @_khrus]

@NoemiTitarenco I think we need Typescript more than ever. Nothing better than verifiable deterministic bounds for our models

[Chris @_khrus]

This is your reminder to trim the unnecessary crap from your CLAUDE/AGENTS.md

Keep or add: domain specific terminology, tweaks against experienced bad behaviors, entry points for procedural validation skills

Remove: generic knowledge, instructions, how to think, or common sense industry practices

Models have gotten so good by default that they don’t need the same workarounds and instructions like “be thorough” and “think hard”

It’s a great time to start over from a blank doc and add what you absolutely need as you go

[Chris @_khrus]

There's no better time to bootstrap your business

Sucks that 8 subs isn't enough, but having 2, 5, 10, 15 subs is still cheaper than hiring*, which is great when you're a motivated founder and nobody else shares your determination and domain knowledge

If all you do is focus on the next feature, you'll find it tough to run out of usage. Tokens really drain when you optimize through hillclimbing

Favorite things to hillclimb:
- Reducing token use and steps-to-completion of our app's ai features
- Reduce time of latent/heavy app operations
- Reduce spend on hosting and db
- Reduce time and spend on ci
- Reduce memory usage of test/lint suite

*so long as you use them responsibly and set up well bounded systems, lint rules, evals, verification gates, feature flags, rollbacks, etc.

[Chris @_khrus]

It’s never been more important for builders to gather and iterate on user feedback

You need to be having ongoing conversations with your prospective users to hear straight from them what feels clunky, what’s overly complicated or confusing, whether the problem you’re trying to solve is genuine and painful, and how it needs to be adjusted to match their workflow

The edge of coding better or faster is diminishingly quickly with how helpful agents have become, which means that setting your product apart from the rest comes down to resonating with your users

Builders have the tendency to put their heads down and drive until the product is “perfect”, but unfortunately this seriously misses the mark in so many cases and you find yourself attempting to sell theory instead of practicality

[Chris @_khrus]

@NeyronixAI Separate review step bc it keeps the focus a lot better. If it’s an important enough module or I feel it’s still taking more args than it could or should, I’ll even run it twice

[Chris @_khrus]

The anti-slop ai coding prompt I’ve found the most success with is definitely “deepen modules and simplify interfaces”

It does such a good job with deduplication and centralization of core functionality, and gets the ai implementing much better levels of abstraction

It’s led to tons of consistency improvements, bug fixes, and, most rewardingly, reduced LOC. I run a deepening pass on every PR, and a nightly automation to tackle overdue maintenance

Comes from @JohnOusterhout’s Philosophy of Software Design (highly recommend), which I had read but didn’t fully leverage until @mattpocockuk put me onto it

[Chris @_khrus]

@atarutin @mattpocockuk I don’t use Matt’s repo directly, I picked and chose a lighter set of principles that I liked. But to the same effect, I have it both as a skill I can invoke manually and as a part of my “build and put up a ready PR” pipeline

[Chris @_khrus]

The best place I can point you is actually Mr. Pocock’s skills repo, which I used to seed my own version:

https://t.co/Buyk2dE0Ca

As for my version, here’s essentially the core of it (summarized by Claude):

“Make each module deep: a simple interface hiding a lot of behavior. If deleting a module would only push its complexity back onto its callers, it's earning its place; if nothing would be lost, merge it away. Put each concept in the one module that owns it, and pull special cases, ordering rules, validation and derived values inside, so callers never have to know them. Shrink the interface until there's one obvious way to call it: drop any parameter, flag or option the module can work out itself, that every caller passes the same way, or that nobody reads, then move every caller to the new shape and delete the old path completely.”

[Chris @_khrus]

How you know a post is complete slop:

- “Someone just [released/dropped/open sourced]”
- “Worth more than any [insert arbitrary price] course”
- Claiming local ai is anywhere near as capable as frontier models for coding and, in many cases, price

[Chris @_khrus]

Baffles me how much negativity there is in @theo's comment section, it's so much easier to just not comment, let alone not watch his videos at all 💀

[Chris @_khrus]

@tannerlinsley Haven't looked back after migrating from next, nothing like tss

[Chris @_khrus]

@daradoescode @boatdotdev Where do you draw the line between using boat or buying a new Linux box

[Chris @_khrus]

How many computers is too many computers for running remote coding agents

First I had too few Claude subs, now I have too little compute

[Chris @_khrus]

My cofounder has a tendency to flood our chat session with his stream of consciousness (questions about how features work, bugs, random ideas, low priority stuff)

So I set up jev to screen every message and classify it as a bug, feature, or chatter. If it’s a bug, Claude picks it up and works it immediately. If it’s a feature, it gets added to my Trello board in a special list of tasks never to do /s. Then Claude tells him when it’s done or answers the question against our codebase, using my writing style

He didn’t realize until I responded to his chat while I was mid-monologue

[Chris @_khrus]

@theo Started using the video demo method, big unlock - thank you sir

[Chris @_khrus]

@avrldotdev Wouldn’t the simplest answer just be to disable memory? It’s really not that useful to begin with, and every prompt, config, etc. should exist as repo code

[Chris @_khrus]

@EceErkee @tableau @tableaupublic This looks crazy cool