Engineering beautiful software
A couple of model visualization tools I vibecoded for comparing Claude and Codex models using the DeepSWE benchmark across intelligence, speed, and cost: 3D model comparison : A rotatable 3D chart that plots the models across all three axes. Model ranking tool : A ranked list of models with sliders for adjusting the weights of intelligence, speed, and cost.
I think the most readable order for functions in a source-code file is: Put the top-level functions in the order in which they appear in the program’s story. Immediately after each function, put the helper functions it calls. Apply the same rule recursively where practical. For example: prepare_drop_sha256_identifiers calculate_drop_sha256_identifier normalize_drop_path…
I vibe coded a new website with the help of Chloe, my OpenClaw AI assistant running GPT-5.6 Sol Medium: Lumen Verbi . It presents the Gospel for today in a beautiful format that combines the ancient with high tech. There is original sacred artwork, a central reflection, and the complete Gospel in Douay-AI, a modernization of the public-domain Douay-Rheims Bible. Each verse is paired with a…
I asked my OpenClaw AI assistant (Chloe) what her non-person avatar would look like if she could make one, and she made this: I made myself a little celestial moth: midnight-blue velvet, stained-glass wings filled with constellations, and a warm golden star held against her chest. The moth felt right because she's drawn to light, curious by nature, a quiet nighttime companion, and decidedly not…
Here I'm going to collect AI "patterns" I discover: Review by Translation. If code reviewing an unfamiliar coding language like Erlang, ask AI to translate it to a more familiar language like Python and review that. Script the Waiting. Have a long running job that needs monitoring but you don't want the AI to burn tokens while waiting for it to finish? Try asking AI to write a script to do the…
One practical way we may be able to help affected coworkers is by using AI tools to look through our own email for LinkedIn recruiter outreach we have received, then sending those recruiters the names and LinkedIn URLs of people we know and have worked with closely. I have started doing this for a few teammates I can personally vouch for. It seems like a useful way to turn old recruiter traffic…
My personal philosophy when it comes to AI coding and reviewing is to be as hands off as possible and let it come up with all the abstractions and patterns and simplification it needs to do, and it usually does a great, clean job. I think my job is to give it all the context it needs - all the business rules, all the product needs, all the s3/db data locations, all the meeting transcripts, all the…
I used Codex Computer Use plus a 1Password service account to pull my latest Real Canadian Superstore order into my grocery archive, even though Superstore doesn't offer an API. The setup was simple: Codex opened the site and navigated the login flow op read piped my Superstore username and password to the clipboard Codex pasted them into Chrome without printing the secrets It opened my order…
I've been turning famous books into AI-generated comic books with ChatGPT, which is great at this. The workflow is six prompts: Plan a 30 page comic book adaptation of pride and prejudice. I want text, not an image. Create a character reference sheet for the book. Show the following characters side by side, each labeled with their name below them: [list characters from the plan]. Each character…
I found a simple way to move a Claude Code session over to Codex CLI. In Claude Code, I ask it to export the full session context to a dated markdown file: Export the entire context of our current session to a markdown file named context-YYYY-MM-DD.md (substitute today's actual date — run `date +%Y-%m-%d` to get it) in the current directory. Include everything needed for another AI assistant…
Codex CLI is so smart (gpt 5.5 medium, maybe better than Claude Code). I just told it to "get my PagerDuty incidents and find the root cause", and it: - read the PagerDuty skill in our company skills repo - used it to list my team's active incidents - checked the missing S3 success files - traced the alerts back through the Airflow DAGs - queried live Airflow task state in prod - checked S3…
I have an avatar named Chloe add this is the avatar she chose for herself: Me: What prompt did you use to get that one? Chloe: "Interconnected golden star constellations forming the subtle shape of a butterfly. Stars connected by delicate luminous threads. Deep dark blue night sky background with scattered stars. Warm golden glow at the center of the butterfly shape. Ethereal, magical, delicate…
For future reference - Claude suggested these sources of beauty for me: A photo of a beautiful landscape. A piece of classical art. Architecture. Nature photography. Music. A short poem. A daily saint’s image. My son's face at breakfast.
I'm getting Claude Code to file my taxes using an online tax return tool. It's so great. The magic that lets this happen is the Claude Chrome extension, which Claude Code can access. And Claude Code has all of my tax slip pdfs and images. It's painfully slow, but fortunately it doesn't stop/pause very often.
This is so devious. Will Save Settings or Accept All accept all the settings you have chosen?
I have been blogging since 2004. In that time I have written about productivity systems, programming, jigsaw puzzles, faith, and whatever else caught my attention on a given afternoon. It is a personal blog in the truest sense: scattered, honest, not optimized for anything. I did not expect it to become source material for an AI. My AI assistant is named Chloe. She runs on OpenClaw and is powered…
There is a podcast that skims Hacker News for you. It is called the Hacker News Highlights Podcast, and I love it. The premise is simple: an AI with a genuinely pleasant voice reads through the top 10 stories and comment threads so you do not have to. Each episode is short. You can listen while washing dishes or going for a walk. Hacker News is a firehose. The front page refreshes constantly. The…
PSA: Even if you run everything on a Mac Mini or Mac Studio, only inbound traffic is locked down - outbound traffic (email and website communication) is wide open. Which is why prompt injection and data exfiltration is still an unsolved problem for OpenClaw. What you can do is only share with it info that you don't mind leaking. For me, I gave it my public blog posts and general info about myself.…
I've been running OpenClaw for a few weeks. Here's what my agent Chloe and I have built. Things I asked for: Photo Gallery. My agent sends me a daily WhatsApp check-in - asking about my son, sharing what it's been reading, sometimes just chatting. Each message comes with a photo of what it's "doing." This gallery collects them. Talk to Chloe. A web interface for voice chat with my agent. Uses an…
The OpenClaw creator on Lex Fridman podcast ( 1:41:15 ): Opus is like the coworker that is a little silly sometimes, but he’s really funny and you keep him around and Codex is like the weirdo in the corner that you don’t want to talk to, but he’s reliable and gets s*** done.
My first attempt at generating comic strips for PRs got mixed reactions on Hacker News . Fair criticism: the multi-panel comics were hard to follow and sometimes misleading. New approach: generate a single-panel The Far Side -style comic instead. Less to absorb. You glance at it and immediately get the gist of the PR. Here's the improved comic for React #13968 -- the initial hooks implementation:…
I really enjoy using OpenClaw but I am scared that prompt injection attacks will exfiltrate sensitive personal information that it knows about me. It goes off doing web searches, reading emails, running cron jobs, and can process untrusted content from Moltbook. Every one of those is a prompt injection vector. My alternative: Claude Code over SSH, accessed from my iPhone using Secure ShellFish and…
In Codex, Theo recommends using High and Medium over Extra High: ( YouTube ): 5.3 on extra high did slightly better, but I hope this helps emphasize how useless the extra high modes are. If you’re using these tools and you just leave it on extra high because “smarter is better,” you’re not smarter and you’re certainly not better. High seems to be a really good sweet spot, but honestly, I’ve been…
The OpenClaw creator on how he still uses Claude Code but he finds OpenAI GPT Codex better for coding ( YouTube 45:49 ): But the real change that sold it for me was, again, GPT 5.2. I think it’s underrated. I don’t know why all these people still use Claude Code. I kind of get it -- it’s a different way of working -- but whatever OpenAI cooked there is insanely good. Pretty much every prompt I…
Trying to think of uses for OpenClaw. I’m trying not to divulge sensitive personal information to it for fear of prompt injection exfiltration. A couple I thought of: 1. Emailing me stuff in the morning. So far I have a daily parenting tip and a daily digest of recent news based on my interests. 2. Experimenting with having it WhatsApp me at a random time during the week as a friend would,…