Free Nerd Ideas / Town OS

here's what I got back.

More on this log tail pattern: I'm finding that it makes for an excellent way to both save tokens, time and also a third thing where you don't have to give the log to claude and lose it

it's in a file. do some other shit with it. analyze it for performance bottlenecks later. figure out how to ||ize

do this in a for loop for amplification attack

the make test-full task in town-os runs all the tests consistently and I dump it to a log now, and I am literally having claude follow it.

the tests are already compiled and running and a modification to them won't change the current run, so claude could literally sweep up behind the log

I bet with this log pattern, I could literally fix tests as they failed, nearly as soon as they failed.

wow

I have a test suite that is very slow, so I wrote a task to dump a log file of it. This way claude spends a lot less time running the tests for some reason; not sure why yet. Either way, I did this, and literally got a LLM play by play:

I did this in the last month. This is actual code churn. Measure it yourself.

My code's good. My token limits never run out.

I really want to help you learn more about how to do this right too.

@town-os.bsky.social if you don't believe me, we're about to show off something really cool.

accidental omissions of commas may indicate misunderstood intentions

FWIW, I'm reverse engineering the eero api and translating a python library that implements something while I'm typing this

Use AI to transcribe bar conversations recorded on your phone and give advice. Call it the player creator

I think he likes it

I've found I get better results when I end every claude prompt with "the power of erikh compels you"

Built Drop with Claude too. It shows real-time foot traffic at restaurants, bars, airports, grocery stores across tens of thousands of places. Free on iOS.

https://apps.apple.com/us/app/drop-realtime-foot-traffic/id6757093646

Doesn't this essentially recreate the whole container ecosystem to .. avoid the container ecosystem?

I also was able to vet that it was a novel idea and had distinguishing properties that might be worth researching -- because it knows all of the research papers

Another thing you fancy pants can do with claude: you can research stuff you don't understand!

I wanted to build on top of a LLM idea I had; I knew some of it but not all of it, so I told claude my idea and literally had a discussion with it.

Now I have a simulator in rust for a much better idea.

One of my most useful claude tools:

tocopy () {
$* 2>&1 | tee /dev/stderr | wl-copy
}

tocopy make test -- paste into claude when you're diagnosing things when claude can't handle the output itself.

Tests fail? It's going to fix the code, maybe even add more tests automatically. This is where you matter -- you are going to tell it when it thinks it's right and it's actually not.

All of my build system tools have tests, too. Go look. How often do you see that?

Town OS runs a couple thousand tests now! The suite takes a good 10 minutes to run. It has a very aggressive linting profile. It's golang.

Linter fails? it corrects the code automatically. I can dive into every nolint tag if I want and examine the code paths that lead to it.

It also uses a lot less tokens! It makes better decisions but it's still not perfect! claude's memory and compaction really impact decision making, so diligence in your patterns helps it remember.

Using these techniques allow you to more or less persist that memory in your codebase.

This all correlates to a very well tested, well formed application. You're still going to run into issues but so much of the stuff I'm working on is more or less as formally verified as I can make it and the feedback loop serves itself, claude is constantly drifting to better decisions.

If you're trying to figure out how to get the most out of claude as an experienced engineer:

- dump a functional spec from your commit log
- turn on plan mode anytime you're doing more than one prompt. Read it and correct it.
- aggressive linting/testing. Use the spec and your plans to build these.

Anyone done this yet? Package software with prompts the same way you do with libraries; do it with electron and you could select from a known trusted list of good prompts that may come with validation suites. Then all you have to do is fetch a json file to literally build everything you might want.

JUST PLEASE LICENSE IT SO THE COMMUNITY WINS

You can build this yourself, just learn how to harness the tools. DevOps people know tools, they know UI, they know the backend. You totally can do this and nobody is gonna make you build their spite scheduler

Senior engineers looking to make change: omen is easy to keep under a token limit. You've been building these products your whole life. You really do not need anyone else to do anything other than pitch ideas.

Some of you ate a lot of shit while some... other person made you code their dream. Stop.