d-lo

I’m very conceptually interested in AI assisted individual learning; I’ve managed to get into a lot of topics that I would have had a hard time approaching without an LLM helping. I really like having AI rephrase concepts usually visualizations and analogies that resonate with me personally. That said, I think there is a lot of sketchiness around Alpha Schools in particular: https://buildcognitiveresonance.substack.com/p/special-repor...

There's another good write up and discussion on Reddit (with fewer LLM writing ticks): https://old.reddit.com/r/ClaudeCode/comments/1ujilqt/anthrop...

I do wonder if Anthropic is modifying output via prompt modification or some of the Fable style weights adjustments for requests that contain these sentinel values. That would be one way to try to prevent distillation, and they have shown a willingness to silently modify model behavior for user input they deem dangerous.

I was largely inspired by beads (https://github.com/gastownhall/beads) but ended up vibe coding something slightly more tailored for the way I like to work. It’s basically a JIRA style task with a few fields (project, title, status, description, priority, parent task, and an array of notes) that I store in a Postgres database hosted in Neon. I have a lightweight API and human readable UI hosted in fly.io. The CLI is a Python API client that gets executed with uvx and can retrieve a list of tasks in JSONL format, get task details for a specific task ID, create a new task or update an existing task. I defined a skill for Claude Code and Codex to use the CLI and was off to the races- both harnesses are very good at following instructions to use the task tracker to manage work items.

This approach works well for me because I can easily view the task list and notes from any device and coding agents can coordinate across repos or development environments. In hindsight it might have been easier to wrap something like Todoist with a CLI so I didn’t have to host any of my own infra. I do think writing your own task manager with a coding agents is almost a standard right of passage in 2026- if you’re going to do that you might as well make one that you can use to coordinate work with LLM coding tools.

I wouldn’t say that my prompts are more proper than anyone else’s, but GPT-5.4 adheres to them really well and does a good job of asking clarifying questions when needed. Spending some time up front to make a very detailed list of requirements here really helped. I think Claude Code still creates better UIs when given a broad task, but Codex makes totally usable interfaces when it’s fed designs and specific instructions. Your mileage may vary; I think different people’s development styles may mesh better with different models and harnesses. I don’t think that Claude Code is a clear winner in April 2026 compared to how it outperformed other options a few months ago.

I’ve mostly switched to Codex (GPT-5.4 high) over Claude Code (Opus 4.6) in the last few weeks. I think my style of workflow and prompting seems to generate a bit better results with it. I have a thin CLI based issue tracker and an associated skill that I use instead of markdown files and Codex does a better job of interacting with it and updating existing tasks over time whereas Claude Code tends to create duplicate tasks with slight changes. I find the code quality of both to be pretty on par with some style differences; Claude has been leaving very verbose comments as of late

I do slightly prefer the Codex app over the Claude Code TUI and Visual Studio Code integrations.

I’d love a federated Strava replacement. Unfortunately I don’t believe that the AT Protocol supports private or limited visibility posts yet, which I think is a pretty key feature for Strava’s use case.

Also a big fan of directly running my projects on VPCs. I do a lot of development on my Linux machine so I really enjoy not having any substantial differences between my local and deployed environments.