jbdamask

Steve's mega-project this past year has been Gas Town. It's a way to manage umpteen coding agents doing your bidding. It allows humans to delegate more to AI while not sacrificing quality. I finally found a day to dig into it and, even though it did some amazing stuff, I was confused.

Gas Town has its own vocabulary since Steve got to name everything. There's the Mayor, Rigs, Polecats, Witnesses...zoiks. He blogged a lot about Gas Town but it's still the kind of thing you need to use in order to grasp.

To make learning Gas Town a bit easier, I wrote a post called A Random Walk Through Gas Town. Well, I didn't actually write it. I indexed the codebase then had Claude pull out interesting features and write about them from the Mayor's quirky perspective.

Feeling great. It’s never been faster to bring ideas to life. If anything, I need to look at nature more

Previously: https://news.ycombinator.com/item?id=47195123 (March 2026, 136 comments)

Several months ago I posted about my little proto-app, Now I Get It! that uses LLMs to translate scientific papers into web pages that anyone can understand. The feedback and engagement from this community was terrific and helped me turn it into a real product. It's a niche, convenience product. Not a game changer. I'm comfortable with that and went into it with eyes wide open.

Regardless of whether the product does anything, building and launching this as a solopreneur was super fun and educational. It helped me hone my agentic engineering skills and also gave me a better sense of what I could truly delegate to coding agents and where I had to be involved.

In short, my experience led me to believe that the best models, harnesses, and techniques aren't yet there to fully build a cloud-native app (AWS) that is fast, cost-efficient, robust, and secure. Not to mention modular and extensible. The number of decisions a human must make is still very high. That's not a huge surprise to me since coding is just one part of engineering.

The fun part of engineering (for me at least) is putting shape to a problem and thinking at a systems level even when decomposing into tasks. It's the constant zoom-in, zoom-out interplay that first attracted me to software engineering. Yes, I rely heavily on LLMs for development but 9/10 times I would find something wrong in a plan, a design, an implementation, or a test that would have created problems if I wasn't there.

The process I followed for this build, in-a-nutshell: 1. Every feature, bug, and improvement gets its own Issue. Descriptions are sparse so the LLM can really dig into during planning. 2. Plan every issue but override Claude Code's planning mode since it tries to implement immediately after a plan is approved. Also, it's helpful to have the LLM write a user story before writing the plan so that UX stays front-and-center. I had so many issues where Claude would be like, "Oh. You mean you don't want the user to see a 403?" 3. Read the plan. 4. Decompose into Beads (Thanks Steve). 5. Open a new Claude Code session in TMUX and say, switch to a worktree for issue xyz and build it. 6. Deploy to a test AWS environment. Use Claude for Chrome to do a first-pass check for front-end stuff then I'd follow up. 7. Devlog the work. 8. Make a PR, review, merge, deploy to prod

Three things I did early on that turned out to be very useful: 1. The app was free when I first posted and people from around the world processed a lot of papers. This gave me a TON of data I could analyze to figure out a pricing structure that maintains a sustainable margin. This is important because the app is self-funded; I don't want to have to decommission it because I miscalculated my expenses.

2. Optimized architecture for low idle cost. I don't want a bunch of infrastructure burning a hole in my bank account if I can't get enough customers to sustain the app. This app relies heavily on static html pages on S3 behind CloudFront, along with Lambda, and DynamoDB. All very cost-effective.

3. In addition to my DevLog skill, I made an app-specific skill called now-i-blog-it. This reviews my DevLog for recent updates and writes short, technical blog posts. Of course, LLMs aren't great writers so I wound up with another skill called cringelinter that can be run in a Ralph loop to scrub all the mic-drops, bro-speak, and cringe words from the blogs.

The blog has over 60 entries so far covering everything from architecture, to adding an inline HTML and KaTeX editor, to my first hacker, to launch. Some of you may find it interesting. It demonstrably shows that that reports of the death of software engineering have been greatly exaggerated.

Thanks again to this community for engaging with my original post. It's been a fun build.

Nice job - I'm a fan. Makes it easy to get the big picture so I know where to dive in.

Added a waitlist to the site https://nowigetit.us

I should probably just make a new post but if anyone is still watching, I've made a couple of changes and have more coming:

1) Modes. You can now choose between Native, Non-Technical, or Kid-Friendly mode. Here are results for the same paper Native: https://test.nowigetit.us/pages/62ef3639-02f0-472b-b0a6-79f6... Non-technical: https://test.nowigetit.us/pages/eba32173-269f-4b5c-9df1-c782... Kid-friendly: https://nowigetit.us/pages/3c413c94-d843-4e84-8805-22af16021...

2) I'm working on user accounts with private galleries

3) Getting closer to opening this up with a cost-plus pricing model. Hopefully by next week but we'll see.

Some commenters have suggested turning this into a sort of graph where references are either brought into the page or new pages are made for references then linked. Is that what you're thinking about?

That's a really cool idea. Can you elaborate?

A few people uploaded the Bitcoin paper and I noticed a bug in one where the page just kind of ended halfway through. This was due to my stringent security protections against prompt injection and outside links but I was blocking some legit CDNs, like Chart.js. That's been adjusted.

Thank you. Great to hear!

Totally agree. At the same time, I find that my brain learns best when I ingest the same information in different ways. This app doesn't replace papers; it complements them. Unless you're my mom - she's not going to read arXiv anytime soon.

No. At least that's not what I intended.

I find the same thing - that sometimes less is more when it comes to prompts. Especially when the inputs are somewhat unpredictable.

Yea, I was surprised by the output myself. It's all auto-generated.

I'm considering some ways to direct the LLM but we're in this funny period where models are getting better on subjective things like look-and-feel. And if I direct too much, I may wind up over-fitting for today's models.

Thanks for the question and links.

I haven't done any token/cost optimization so far because a) the app works well-enough for me, personally; b) I need more data to understand the areas to optimize.

Most likely, I'd start with quality optimizations that matter to users. Things to make people happier with the results.

I haven't but am open to ideas. What kind of experience would be useful to you?

If by "auto-generated" you mean, does the LLM generate the output from the input, then yes.

I've upped today's (3/1) cap to 100 papers

Doh! I didn’t think of that. Interesting idea.

Very nice of you to say.

Nice

Thanks to everyone who tried this today and those who provided feedback. I really appreciate your time. Here are some stats:

100 papers processed.

Cost breakdown:

LLM cost $64

AWS cost $0.0003

Claude's editorial comment about this breakdown, "For context, the Anthropic API cost ($63.32) is roughly 200,000x the AWS infrastructure cost. The AWS bill is a rounding error compared to the LLM spend."

Category breakdown:

Computer and Information Sciences 41%

Biological and Biomedical Sciences 15%

Health Sciences 7%

Mathematics and Statistics 5%

Geosciences, Atmospheric, and Ocean Sciences 5%

Physical Sciences 5%

Other 22%

There were a handful of errors due to papers >100 pages. If there were others, I didn't see them (but please let me know).

I'd be interested in hearing from people, what's one thing you would change/add/remove from this app?

That decision-tree page is killer!

Sorry you hit this. 100 papers were processed today. Cost to me was $63.

Very cool. Appreciate it.

Agree. Curious if you’ve played with landing.ai?

I hear you.

For me personally, the pain point is being interested in more papers than I can consume so I’ve gotten into the habit of loading papers into LLMs as a way to quickly triage. This app is an extension of my own habit.

I also have friends without scientific backgrounds who are interested in topics of research papers but can’t understand them. The reason for the cutesy name, Now I Get It!, is because the prompt steers the response to a layperson

Hmmm, didn’t realize I was deflecting - just stating facts. But if I came across that way then criticism noted.

If I turned this into a paid app then more attention would be given to quality. There’s only so much an app that leverages LLMs can do, though. With enough trace data and user feedback I could imagine building out Evals from failure modes.

I can think of a few ways to provide a better UX. One is already built-in - there’s a “Recreate” button the original uploader can click if they don’t like the result.

Things could get pretty sophisticated after that, such as letting the user tweak the prompt, allowing for section-by-section re-dos, changing models, or even supporting manual edits.

From a commercial product perspective, it’s interesting to think about the cost/benefit of building around the current limits of LLMs vs building for an experience and betting the models will get better. The question is where to draw the line and where to devote cycles. Something worthy of its own thread.

Sorry. Reached 100 uploads today. Check out the gallery

LLMs, even the best ones, are still hit or miss wrt quality. Constantly improving, though.

I see more confusion from Opus 4.x about how to weight the different parts of a paper in terms of importance than I see hallucinations of flat out incorrect stuff. But these things still happen.