Thread

This is the way 🤙

Awesome, nice work.

👀

In case this saves you some time (I'd already made these notes for myself in Trilium): Calle's way Calle's guide to vibe coding (Citadel Dispatch 171) Use different AIs, one for making the implementation plan + requirements, the other for coding up. He uses Goose & Cursor (shows him to switch between models - use ppq.ai for same functionality) If wanting to understand existing code, he used Gemini to explore and describe it to him. His work flow: 1) create Implementation plan (list of things that need to be done in order). Steps to achieve overall project 2) create requirements in each step (to fulfil that step) [my addition, after 1st application of this workflow, getting the first AI to produce a simple user manual is also helpful for the next AI & to sanity check that it understands how you want to use your project] 3) pass all of that to a new AI (e.g. Claude) For bitchat, he spent ½ day iterating like this and had a working MVP with basic features. As project scales, or for niche / specialist code (not many examples for AI to learn from) manual coding is required. When project becomes large, AI loses context, large bodies of redundant / unused code etc.

🤝

MVP 2 is live!! 🥳 I'm finding Gemini 2.5 Pro much more productive than Claude Sonnet 4.5 (Claude got MVP1 running v quiickly, but was then dreadful at updates/iteration from there). Now I have a GUI, the ability to tidy up unwanted songs (on the mobile) using mirroring, and the GUI even 'remembers' the last selection of playlists. 🙂🙂🙂 I've learned to use git for version control and have a local repo (next step is connecting to / subscribing to a remote repo). My top tip is, ask the AI to freeze the interfaces as much as possible. Before I did that we were chasing our own tails. A bug fix would be implemented but also include name changes to keys or functions, thus trigerring more debugging which changed more names. After asking to freeze them, this (almost completely) disappeared. 😏 Watch out for the AI amnesia episodes. For a tool that is impressive at extracting meaning from my instructions and then pretty autonomous delivery based on those instructions, it often forgets why it did something 2 steps ago. I'm frequently asking, 'haven't we just undone that feature that you added because <reasons>!' So many thanks to nostr:nprofile1qqs9pk20ctv9srrg9vr354p03v0rrgsqkpggh2u45va77zz4mu5p6ccpzemhxue69uhk2er9dchxummnw3ezumrpdejz7qgkwaehxw309a5xjum59ehx7um5wghxcctwvshsfmrzrk for sharing his way of vive coding & nostr:nprofile1qqsdy27dk8f9qk7qvrm94pkdtus9xtk970jpcp4w48k6cw0khfm06mspp4mhxue69uhkummn9ekx7mqpz4mhxue69uhkummnw3ezummcw3ezuer9wchsj08ren for making it so easy to try out different models, cheaply AND WITHOUT KYC gate keeping / control. Bless you all. Now to find a harder challenge...

Holy sh#t, IT WORKS!! My coding AI & I have reached MVP3, I can now sync the music to either stock android or to a GrapheneOS device (and THAT, dear reader, was a painful journey, the last *day* of which was down to a missing '/' 🤦). FWIW my feedback on vibe coding models (after about 3 weekends) is that Gemini Pro 2.5 seems much more useful than Claude Sonnet 4.5 after you get past the initial MVP. It is better at debug and iteration (IMO). Oh and it's also MUCH CHEAPER. nostr:nprofile1qqsdy27dk8f9qk7qvrm94pkdtus9xtk970jpcp4w48k6cw0khfm06mspp4mhxue69uhkummn9ekx7mqpz4mhxue69uhkummnw3ezummcw3ezuer9wchsj08ren already makes it easy, relatively cheap (see below) & kyc free to use any model you want (🙏). It's also really easy to see how quickly each model uses your credit. In the debug/iteration phase, I found Claude Sonnet 4.5 was burning through $10 in a weeknd EASILY. Gemini Pro 2.5, by contrast, burns about $2-3 in a heavy day ehich is largely (ok, antisocially) dedicated to vibe coding 😶‍🌫️ Time for a celebratory dram. 🍻