Mark Dalgleish

We wrote a blog post that explains the new @ReactRouter RSC architecture a bit more, and how it makes our library APIs almost as powerful as Framework Mode. remix.run/blog/react-rou…

Ladies and gentlemen I present you a @markdalgleish twitter post screenshotted to reddit then posted back to twitter - attributed to the reddit page.

@Sc_Meerkat This is what I said when I had to maintain the tests by hand, but this is exactly what’s changed.

@markdalgleish The tests are never free. There is a cost associated with modifying tests when modifying code. When you test internals the likelihood of tests requiring changes is higher.

I used to be against a lot of heavy unit testing of internals. I felt they often cost more in dev time than they were worth compared to e2e. Now that AI makes these tests effectively free, and agents use them for localised feedback, this equation has shifted back a lot for me.

“Skipping tests to get a green run was the wrong approach.” Sometimes you've gotta be patient with the little guy.

@royalicing Exactly. I used to feel like a lot of these tests cost too much to maintain, but now the agent knows when to update or discard tests and can clean up as it goes.

@markdalgleish Do you keep the tests once it’s done? I guess your point is it’s fast at rewriting tests too, so they don’t have to become a burden.

@royalicing So agents know their code works at every layer, and have a good place to diagnose and fix issues in isolation. Without this the dev loop is too slow.

@markdalgleish Are you checking that these tests do what you expect? Or is it more for the agents to stay aligned? I still think testing at the significant boundaries is what I most value. Especially because they’d be finite enough you can review them.

Coding with agents is so addictive, it’s surprisingly easy to lose track of time. I feel like there’s an extra layer of productivity gains in the fact that I can get into a flow state almost instantly, and stay there much longer.

@ScriptedAlchemy Man, you’re always on another level. Do you have more detail on this? What do these ai.tsx files contain and what’s the workflow around it?

I switched to the module bundler style where I register a new parser and module factory on the bundler for “ai.tsx” modules. Which use a special parser that build hooks are agents. Dependency tree and chunk graph naturally take care of context splitting management and agent orchestration as a side effect of expressing the prompt

The biggest takeaway from the whole “Ralph loop” thing for me so far is realising that I’ve been using Markdown “plans” with the agent too much when this isn’t my natural workflow. I’ve tried switching to a Kanban board inspired flow using Markdown files and it’s so much better.

@taayjuss So far I just add a task to explore an area further with my rough thoughts so far.

@markdalgleish how do you handle the tasks that aren't clean enough for kanban yet? or do you just force everything into cards

@_aierie I haven’t used it, just learning from it. I wouldn’t feel comfortable letting it rip on a massive list of tasks at the moment. I want that feedback loop of being able to play around with it at each checkpoint because that’s part of the design process to me.

@markdalgleish What have you used Ralph for so far that you like?

Right now I still have a plan file, but it kinda morphs into more of a living architecture doc while iterating. The actual work I’ve got structured like this: /tasks/*/todo.md /tasks/*/in-progress.md /tasks/*/done/001-task-summary.md /tasks/*/done/002-task-summary.md

Claude: Hey, mind if I grep -ohP "useEffect\(.*?\[\K[^\]]+" **/*.tsx 2>&1|tr ',' '\n'|awk 'NF{$1=$1;a[$0]++}END{for(k in a)print a[k],k}'|sort -rn|head -20 Me: ... yeah go for it dude

@scalinglaws @pangramlabs It does sound very AI generated now that you point it out.

@markdalgleish @pangramlabs i was curious about the two "it's not A, but B" in your second point; guess i'm overly sensitive 🫡

Absolutely mind boggling to stop and realise: 1) I barely code anymore, just chat with an AI. 2) It’s not slop. I’m still engineering. I don’t feel threatened. Feels like pair programming. 3) I’m enjoying it more than coding by hand. Truly wild time to be living through.

@pangramlabs @scalinglaws I did actually write this 100% by hand on my phone.

@scalinglaws @markdalgleish We believe that this document is fully human-written Disclaimer: For text under 75 words, results may be less accurate. pangram.com/history/1fee9a…

@kentcdodds @simonswiss @glenmaddern @JedWatson @philnash Merry Christmas!

Merry Christmas to my friends in Australia! @simonswiss @markdalgleish @glenmaddern @JedWatson @philnash

Tired: Copying and pasting their docs into the prompt. Wired: Cloning their entire repo into a gitignored directory and then asking the agent about it.

@ryanflorence @BrooksLybrand I went to add something like this to the Remix repo and realised you’re way ahead of me. Such a great pattern.

@BrooksLybrand @markdalgleish it's lovely

When working with LLMs and new libraries they might not be trained on (like new Remix stuff), just do the following: mkdir references echo "reference/" >> .gitignore cd reference git clone git@github.com:remix-run/remix.git You've got all the context you could possibly need

@mjackson Fun fact, Gatsby was originally built on top of this before they built their own: github.com/markdalgleish/…

Remember React static sites and Gatsby and Jamstack? Ha, that was a funny time.