Mark Sanborn

@thdxr I just have LLMs run on a cron doing small refactors, unit tests, etc all the time.

are there people out there who just want to refactor every day? just wake up and find the worst code and just chip away at it and clean it up wake up the next day do it again, infinitely improving things with zero external impact?

@ThePrimeagen You have to just embrace the slop. You can prompt a refactor, cleanup, unit tests, benchmarks just as easily as you can prompt features. Let go of the code.

I am slowly coming around to AI assisted programming. I am genuinely trying to codify every rule about programming that I have and using that + several stages to build out small changes. Not sure the productivity changes, but I think I can see a modest gain in speed. I am also trying to be concerned about every line produced, not just slop trebucheting code over the wall.

It gets worse. Once you finally figure out how to sign and your signing for a few months you will find a random build stuck signing with no error. All builds will fail for days. You’ll see that the forums are full for people with your same problem. You learn that you have to reach out to support because something in your build triggered human review but it never gets unstuck until you send like 10 support requests and wait a few weeks for some human somewhere to click an approve button.

paid for an apple developer subscription 5 days ago, $100 cleared out of my bank account, and i'm still not approved, and i still can't sign apps. i'd have never thought i would have to chargeback _apple_ of all companies.

@kevinrose On #4 just do /clear. Also you should do this a lot. Models get worse the more context they have.

i'm sure we all have our little coding 'hacks,' here are my top 5, please share yours (or help me improve mine! 🙏): 1. Plan -> Deepen Plan -> Then let Codex review the plan, then hand it back to Claude Code 2. Let Co-Work read the plan and build you a PDF of the plan in plain english along with flowcharts (vs just "go to work!"), this is a great for overall logic agreement 3. If I'm unsure of a stack or an algorithm choice (e.g. best algo for clustering objects with vector embeddings), give it to the beast models and let them deep research it for 20 mins 4. If have something big to tackle, always quit and restart Claude Code 5. On big PRs, I always let CC, Codex, and @greptile view it (at the same time), never fails to find some P1s

I love finding good static prompts I can loop a few times. My latest: “without adding any new features what is the most useful thing we can add to make the user experience a pure delight to use?” So far about 9/10 of the suggestions I’ve been like hell yeah implement it now.

@karpathy Do you keep it in a git repo? Regular backups?

@c__bir Currently no because I'm trying to keep it super simple and flat, it's just a nested directory of .md files and .png files and a few .csv and .py, and the schema is kept up to date in AGENTS.md . The LLMs get this very easily. Any custom functions are easy to vibe code tools for.

LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So: Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them. IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides). Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale. Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base. Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into. Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries. Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows. TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

Just saw Claude code use Perl. That’s awesome to see.

@d4m1n Doesn’t the credits expire in a month?

✋ stop what you are doing and do this instead > go to Claude․ai > settings > usage > claim $100 or $200 in free credits they finally did something about it.

A simple Ralph loop along the lines of “look for code smells and find the most critical refactor or way to simplify this code base or make it more maintainable. Do it while making sure all tests pass. Do not modify test logic.” Does wonders. Run this like 50 times until it starts doing minor stuff and stop running it.

I think it's probably undeniable that RIGHT NOW, if you effectively vibe code your projects, the codebase will become increasingly worse with each passing day. Not sure that any combo of tests + formatters + skills can prevent that.

@GamewithDave Age of Empires

For those who used a computer between 1995 and 2001, what's the computer game from that time that sticks with you the most, and why?

@ScottApogee I got this game from Walmart when I was like 12 and I was disappointed to find out it was censored. Had my 14-15 year old friend call in and pretend to be my mom to get the uncensored patch sent out. Arrived on a 3.5 floppy. Best game ever at the time.

We wanted to start the game off with a fun opening. Did we succeed?

@levelsio This guy ships

This took 24 minutes Back when I did 12 startups in 12 months in 2014, it took me at least a month to make a new startup and I didn't even finish it And they were in fact as basic as this kind of, because they were MVPs too! If 'd ship every day and I sleep 8 hours I have 24-8=16 hours * 60 minutes = 960 minutes per day. And it takes 24 minutes per startup with AI to build an MVP So theoretically you can build 40 ideas per day, and if you work 6 days a week that's 1,000 per month and about 12,000 per year So a decade later, you can now build 12,000 startups in 12 months

I wrote an app in clojure because I wanted to learn lisp back when I was manually writing code. Now that I’m not writing code it doesn’t need to be written in a fun to write language. I ported the entire thing over to rust in less than 24h. LLMs are really good at porting given reference code. The app is way faster and better.

@levelsio Just make sure you have tailscale acls setup right or one hacked server allows attacker to attempt to move laterally on all machines in your tailnet if you just allow everything.

If your Tailscale is hacked The hacker now has direct access to your server But now he still needs to get into your SSH with an SSH key So to get in two extremely rare things have to happen: 1) Tailscale is hacked 2) There's an SSH 0-day that they can use to hack into your server That's still superior to NOT using Tailscale where they'd only need: 1) There's an SSH 0-day that they can use to hack into your server It's like saying "why do you use an alarm system for your house, what if it breaks?"

I’ve been considering this. I’ve already vibe coded a Quickbooks replacement. Claude even helped find a discrepancy. I think that was my aha moment. Not only can it build the accounting software but it can help find mistakes in the data and I can ask accounting questions at the same time.

Fired my accountant. Claude Code is doing my taxes this year.

@unclebobmartin you can add features faster than you can make a jira ticket. 24/7 always on AI QA is a thing now. I used to think AI meant lots of slop. I think it’s the opposite. AIs ability to find bugs, refactor, and do QA is incredible.

I have an itchy feeling that AI development is going to be a boon for QA and Jira.

I just vibe coded a quickbooks replacement for my needs. P&L, balance sheet, auto transaction categorization and CSV import for my bank. Easy to check it did the right thing by comparing P&L output with QB. I was even off by $1.52 and it had to do with two transactions on a date boundary. AI figured out automatically. It’s literally a better solution than what QB provided for me. It’s faster, local, and I can adapt it. No more slow clunky ad filled QB. Also pays for some more tokens.

Is the new purple on Claude code terminal logo a nod to Gerald Sussman’s fez hat?

@r00k I’ve been toying with setting up daily improvement crons. Add a missing unit test, add a critical refactor, simplify something, give me one SEO improvement, etc. I think I’ll try this. I remember you did a code tidying email thing. You could put each one of those on a cron.

Fun little app idea: Opportunist. Each night it makes its best guess as to what you'd like and opens a PR on all your projects. Closes them if you don't merge the next day. Notices which things you merge to self-improve.

@FlorinPop17 Pretty much 100% since late Dec. I don’t plan on writing manual code unless I have to and even then it will be AI assisted.

How much of your code is AI generated nowadays?