Published
Weekend Reading — First roll back, then debug
Lady Laura "Pack of Dogs by John Littleboy"
Tech Stuff
Infisical I'm trying this service as a replacement for Doppler. Seems to have more capabilities plus it's an open-source project you can self-host (if you so care). Does feel overwhelming at first with all its capabilities but I got my training wheels on Doppler so I think I'm ready for this new challenge.
The unwritten laws of software engineering
Here are 7 laws every engineer has broken at least once, learned the hard way:
- It’s always related - first roll back, then debug
- Backups aren’t real until you’ve restored from them
- You’ll always hate yourself for how you write logs
- Always have a rollback plan. ALWAYS.
- Every external dependency will fail
- If there is ANY risk - “4 eyes” rule
- There is nothing more lasting than a temporary fix
You’re still designing for an architecture that no longer exists
Here’s the critical point: these aren’t features of Claude. They are coordinates of a new architecture. Any system that embodies intention, autonomy, adaptation, and orchestration is operating in this new space — regardless of which company built it. Claude happens to be the most complete manifestation today. But the architecture is bigger than any single product.
Will AI Kill Open Source? I hope not, as I'm working on a new open-source project that's focused on AI:
Here’s the workflow that actually works:
- Choose your frameworks and specify your stack up front
- Work in small, reviewable chunks – not overnight agent runs
- Use spec-driven development with tests and architecture decisions
- Review all AI-generated output before shipping
Claude picks the first idea that works. Make it pick the best one. The replan skill that will critique, generate alternatives, develop the best one, and repeat.
Why MCP is Key for Non-CLI Agent Integration and Security
Holmes is writing from the perspective of a developer sitting in a terminal. From that vantage point, everything he says is correct. If your agent is Claude Code or Gemini CLI, running in a shell session on your laptop with your credentials loaded, then yes, gh pr view is faster and more capable than any MCP wrapper around the GitHub API.
Related, MCP-UI to build interactive UI components over MCP.
agent-404 Smart. I don't know how effective it is in the real world, but it's an interesting idea and fairly easy to add:
Claude, Cursor, and Copilot follow URLs baked into training data. When your docs restructure, they get a 404 and hallucinate. agent-404 returns structured JSON-LD so they find the current page instead.
GitAgent Turn your Git repository into an agent because you can version control its memory and soul! I already have a few skills I'm using continously (product planning, finance, marketing, etc) so I decided to turn them into agents. Each agent remembers the decisions I made, things I instruct it to recall for later, has a soul and rules to abide by, etc. And super easy to version them and git push to backup these changes.
Supergood Turn any website into an unofficial API that you can code against. I think I have some use cases for this.
sheklldex A registry of OpenClaw clones, forks, derivatives, and inspired projects.
AI Code Semantic vs pragmatic:
Breaks commonly happen when a semantic function morphs into a pragmatic function for ease, and then other places in the codebase that rely on it end up doing things they didn't intend. To solve this, be explicit when creating a function by naming it instead of by what it does, but by where it's used. The nature of their names should make it clear to other programmers in their names that their behavior is not tightly defined and should not be relied on for the internals to do an exact task, and make debugging regressions from them easier.
Chartle Charts without the annoying BI toolkit. Here I'm asking Chartle for the S&P 500 from 2024 to 2026 and it gathers the data and visualizes it. You can ask for anything or upload your own CSV.
Eye for Design
One of the most confusing but ubiquitous patterns in design is “small text with a background color that may or may not be interactive”. To some that sounds like a button with an identity crisis but to others they might recognize this as a few names: chip, badge, pill, tag, or even lozenge. It’s bad enough that the definition of these things is contested, but what is worse is because they are all generally so visually similar, it makes it hard for the users to know what they are meant to do.
‘Your Frustration Is the Product’ TL;DR
The web is the only medium the world has ever seen where its highest-profile decision makers are people who despise the medium and are trying to drive people away from it.
Optician Sans A free font based on the historical eye charts and optotypes used by opticians world wide. (via mako)
Turn your handwriting into a real font 👍
What's My ΔEOKJND? I'm not color blind but I'm also not very color perceptive. This is a fun, not too difficult, test for your color vision.
Peoples
Dare Obasanjo When I use AI I get 10x the productivity bump on 1/10th of the effort which comes out to about 10% gain:
This is the big challenge for companies in the age of AI coding agents
- Our quality processes are not designed to handle code creation at this pace (10x the code = 10x the bugs)
- Our organizational workflows with their various meetings and approval processes are big drags on capturing these productivity gains, yet are important and exist for valid reasons.
Business Side
The companies that win with AI may not look like companies at all Companies that treat AI as operational optimization are going to miss the real transformation:
The wrong question is this: How can AI make our current company more efficient?
The right question is much more uncomfortable: If we were building this company today, in a world where AI already exists, would we build it like this at all?
EdoStra/Marketing-for-Founders That's quite the collection of practical marketing resources helping you get to the first 10 / 100 / 1000 users. I'm going to try some of these over the next few weeks.
Startup Punditry’s 25 Years of Failure Listen to startup pundits. Not because the pundits are telling you anything important, but because the pundits found a way to reach a wide audience, and isn't that the purpose of your startup? Reach a wide customer base that believes your marketing? aka "Do what they do not what they say":
The whole enterprise has the structure of what the physicist Richard Feynman called a “cargo cult science”: an edifice that mimics the form of science without its substance, deriving rules from anecdotes without establishing underlying causality. Just because a handful of successful startups conducted customer interviews does not mean your startup will succeed if you do too.
Machine Intelligence
Literally yesterday I was working on a sqlite3-based experiment with Opus 4.6. Initial SQL queries were successful but Claude looked at the timing, mentioned that 1s is way too much for a query like this, spent 10 more seconds on the code, and then the (complex) query only took 0.02s.
Sorry if that doesn't fit the very simplistic "all AI bad" narrative but as with any tool, it depends on what you do with it and how you use it.
Assaf I'm tracking bot activity for my app, which has reasonable AI citations — 14 citations, a score of 29.7, and 13,037 bot visits. The visiting AI bots include ChatGPT, OpenAI Search, Perplexity, Claude Bot and friends. So far I got 0 requests for Markdown content. Zero. In theory AI bots may favor Markdown because it's simpler to parse, but in practice they seem to not care.
Insecurity
Do I trust this computer?, my iPhone asks me for what has to be the thousandth time using the same computer logged into the same Apple ID. Enter your passcode, it demands once again, reflecting the hardware, software, and services working together in a way only Apple can deliver.
'StravaLeaks': France's aircraft carrier located in real time by Le Monde through fitness app
Arthur had a profile on the Strava fitness app that anyone could view, because it was set to "public." As a result, the young officer almost instantly revealed the exact location of the French aircraft carrier Charles de Gaulle in the Mediterranean Sea as it traveled northwest of Cyprus, together with its escort, about 100 kilometers off the coast of Turkey, Le Monde found.
Everything Else
Barista: Can I get a name for the order?
Me: looks around leans in whispers ...Banksy.
Remember: The dumbest person you know is being told 'you are absolutely right' by a LLM right now.
GentleMan Gef "Corvus Catus"
To keep costs down, most episodes of Star Trek were filmed on Earth instead of on location in outer space.
I dunno about these special IDEs of March. I like my regular one just fine for coding.
Indie devs, rejoice; I bring you good news.
Bug is over.
What If Your Browser Could Just Ask for Markdown?
Every website you visit today sends you a pile of HTML, CSS, and JavaScript — even if all you wanted was to read a blog post. Your browser downloads layout grids, tracking scripts, cookie banners, and font files before you see a single paragraph of text. For a 2000-word article, the actual content might be 12 KB. Everything around it: 2 MB.
Once again I am heartbroken to remind you that the Dunning-Kruger effect is probably not real. Like Freudian psychology, Hardin's tragedy of the commons and any number of other popular pseudoscientific narratives, it caters to our preconceptions and makes for entertaining, easy to re-tell stories, but it's also... not true.
And - again, I am entirely saddened by this - that means that if we keep using these metaphors we're legitimizing the false ideas behind them.
I'm pleased to share with you the greatest advancement in bottom-sheet technology since the invention of elastic.
Why I email complete strangers
The first time I emailed a stranger, I swear my cursor hovered over Send for a full five minutes.
I had plenty of justifiable reasons to remain hesitant. Not wanting to take up their time, feeling bothersome, worried my question was a silly one... A hundred disparate excuses leading back to the same core: "I’m not enough." That’s the forever curse of low self-esteem. The best and worst-case scenario can never occur, because you’ve rejected yourself first. Who knows how the other is going to respond?
That’s where the rub is, isn’t it? The terrifying unknown.
Two things that make me happy:
Telling my friends I love them. I do, so I let them know. It's so nice to hear someone loves you. And like the man said, what the world needs now is love sweet love.
I also start text messages with a term of endearment. "DARLING! Do you want to get together later?" Or whatever. I like the idea that someone gets a notification that starts "Hey there Sunshine!" or some other pleasant phrase.
Winners of the British Wildlife Photography Awards 2026
