Happy Halloween! It's been a while, but here's my list of "good reads" (As usual, a hat tip to Hacker News, Reddit, Youtube, Spotify, Twitter, Bluesky and my other feeds).
Until next time!
Saturday, October 31, 2026
Fall reading - 2026
Until next time!
Saturday, August 01, 2026
Summer reading - 2026
Some reading before Summer wraps (As usual, a hat tip to Hacker News, Reddit, Youtube, Spotify, Twitter, Bluesky and my other feeds).
Until next time!
Thursday, May 21, 2026
An obligatory post about vibe coding
(An obligatory post about vibe coding - everyone needs to make at least one, they say.)
I've been using my $20 Claude subscription to tinker on some side projects in my personal time.
It's astonishing how much can be done with the basic plan, which makes me wonder how much it will actually cost once the VC subsidies end.
I've also gained quite a bit of insight without going neck deep into multi-agent, spec driven, ralph loop, <insert latest buzzword> and such bleeding edge techniques or tools.
I've stuck to a simple workflow
- It involves Claude code on my Macbook, an architecture.md, a backlog.md and a simple claude.md. I also get the AI to create a more detailed but transient plan-<complex-feature>.md for complex, work in progress features
- I use Claude in the browser for learning, research and brainstorming
I've iterated on my personal workflow over multiple projects. It's been "productive", meaning I've created many experimental projects, including some that may have legs. Some have been purely personal and custom productivity type of tools. I've not really read the code it has produced because all of it was done while I was watching TV or doing chores. Some even in programming languages I have no prior experience with.
Learnings (using Claude as of May 2026):
- Despite having Claude.md and tests it can make subtle mistakes. On occasion
- It leaves docs in an inconsistent state and requires a reminder/instruction to find and fix them
- It forgets to read some of the instructions in Claude.md
- It writes docs in incorrect folders ignoring the designated docs/ folder
- It accidentally deletes features as part of larger changes, esp if those changes cannot be tested (UI)
- It freezes
- (My work related/professional experience is different which I won't speak about here)
- Judgement and Taste absolutely matter
- Identifying problems worthy of solving, brainstorming potential solutions and implementing them in the correct order is really where a good human engineer does really well
- Producing a usable application that is based on the most popular frameworks, on which the AI is trained on is a shockingly easy and delightful experience. Can't say the same about juggling multiple agents and reviewing lots of AI generated PRs from multiple team members though
Here are some of my older experiments: https://github.com/AshwinJay/ai-experiments
Here is one that I may actually use one day: https://github.com/AshwinJay/project-tracker
Until next time.
Subscribe to:
Posts (Atom)