Thinking about frameworks as the building blocks we used in the past. We had really good, popular ones used to build great software. The other side of that - the principles or concepts in such frameworks - had to be re-implemented for every environment or tech stack that the given project required.
One of the core values of really good frameworks was the massive amount of time and effort engineers put in to capture quality. At the same time, a ton of experience went into making those libraries as good as possible, and they encoded numerous issues and use cases that translated into quality software.
My view is that frameworks like this will disappear, but they will be replaced by test cases, guidelines, and invariants captured in those frameworks and shared as instruction packages. For example, terminal emulators. Thanks to libghostty, we have a great framework that encodes all the nuances, invariants, performance optimizations, and qualities that define a great terminal emulator. Distilling that knowledge into instructions (tests, formal proofs, etc.) would let AI agents apply it to any terminal emulator project. A human can describe the products they want to build, and using those guidelines, they can apply them independently of the environment the AI works in.
Installable institutional knowledge.
I love reading books; even more, I love reading multiple books concurrently. Lately, I also enjoy using multiple programming languages. Even, and especially, those that I didn't work with before. With AI, without AI, in any way - just breaking stuff and observing how the program/system reacts to the change. Then fixing it, changing it, throwing it away and starting over. I remember the moment when programming clicked for me - it was when I was introduced to a debugger. The feeling I have now is really comparable to that.
With AI, it's a no-brainer to ask it to create multiple variants, iterate, pick one, and share with colleagues. I used this a couple of times lately - it worked really well with stakeholders who have an idea but don't know how to approach it. It becomes a process where they pick from a number of options, sometimes multiple, and ask for a remix or something, and that kind of funnel leads to pretty good results.
I really enjoy this process. Even before AI, there was so much value in creating low-fidelity mockups, opening Figma or Excalidraw, dropping a couple of screenshots of existing UI, and just copy-and-pasting parts of the UI - like a ransom note. Worked extremely well to communicate the idea.
Thinking about frameworks as the building blocks we used in the past. We had really good, popular ones used to build great software. The other side of that - the principles or concepts in such frameworks - had to be re-implemented for every environment or tech stack that the given project required.
One of the core values of really good frameworks was the massive amount of time and effort engineers put in to capture quality. At the same time, a ton of experience went into making those libraries as good as possible, and they encoded numerous issues and use cases that translated into quality software.
My view is that frameworks like this will disappear, but they will be replaced by test cases, guidelines, and invariants captured in those frameworks and shared as instruction packages. For example, terminal emulators. Thanks to libghostty, we have a great framework that encodes all the nuances, invariants, performance optimizations, and qualities that define a great terminal emulator. Distilling that knowledge into instructions (tests, formal proofs, etc.) would let AI agents apply it to any terminal emulator project. A human can describe the products they want to build, and using those guidelines, they can apply them independently of the environment the AI works in.
Installable institutional knowledge.
Lately, the Backlog status in @linear feels more like an escape hatch from Triage. The Backlog is an ever-growing category, and most of the time, tasks get left behind. Moving a task from Triage to Backlog is just moving it to another holding area. I prefer thinking of tasks in Triage as those that should end with a decision to make the task actionable, cancel it, or mark it as a duplicate.
Awesome article in Quanta Magazine on the Four-Color Theorem. Especially loved the part about the proof that used computers and was not well received. Sounds familiar?
https://t.co/Pl5BQA73f6
There are times when I go outside with the iPad Pro and use the native keyboard or a Bluetooth keyboard to work on long text. After working this way for a couple of hours and then switching back to the laptop, I sometimes reach for the screen, trying to touch it and navigate there. There might be utility in this.
In most cases, tests capture the direction agents will be forced to go in. If a project has strong foundations and the tests were added on purpose, they can be valuable for keeping agents on track (ideally a pretty wide track). And if tests were added randomly because instructions say so, the quality will deteriorate more quickly because agents will have no choice but to build in compliance with those tests. This effectively captures inefficiencies (slop, in other terms).