stopped posting causing iโve thrown my harness away, and learning phase quite over, was a great experiment to learn fast and low level about ai, would 100% recommend anyone to do so
back to build my new project with all learnings, locked in, loving it, grateful for everything ๐
like llms we canโt help but to imagine the next note.
like llms we eval ourselves and fine tune us and after 10 listens we predict better but not yet 100% accurate
after 1000 listens we start to be accurate
when singing itโs so obvious we are predicting machines not so far from llms
when you listen to a song, your mind is always half a sec in advance, predicting the key word, note, drum beat.
even when itโs a first time listen
youโll cover:
- spec driven vs iterative approaches for complex systems that harnesses
- agent loop and observability of agents
- how to test a hybrid system (agents + deterministic code)
- how to manage an agent context
- what models are good for what
- know multiple ways of doing memory, context injection, compaction etc..
- how to make it so agent learn automatically
- so many more things
should write an article tbh
spent last 3 months building harnesses, using openclaw codex claude cursor pi tuis & sdks. open weights w/ comfy ui and llama.ccp.
would def recommend to anyone new to ai if you want to do more than understanding agents but really be in flow with them.
You get good at anything by doing it a lot. Just do things a lot
It's not time spent. It's volume of doing. Make sure the time it takes you to do a single thing is as fast as possible. Your cycle time is sacred
I think @antirez ds4.c is important! I wrote down my thoughts on why I built pi-ds4 and why we need to focus our local model efforts stronger than we do currently. https://t.co/61h4JDHTZL
@mattpocockuk yes with claude code and opus 4.7, always use this + required readings in the handoff (docs/ currently worked on code, plans)
though with codex just letting it, fork + compact
iโve done this but after some weeks they become aggregates that are not very usable, whatโs your strategy so to distill them/ maintain them, remove whatโs outdated etc.. ?
btw thatโs why i keep ultra maintained specs to then ask agents to reduce those backlogs/memory against current specs and past commits
totally agree, I think tomorrow's cofounders will be just that:
one tech deep system expert
one artist
the marketings/ sales/ management will be about expanding the artist vibe via the system
Claude Code eng leader @Nerdi_Yogi on the two profiles she's hiring for now:
1. Creative builders with product sense
2. Deep systems experts (for the hard parts)
btw the goals of such a frontend is not usability, it's reviewability of the underlying system
makes it 10x better to review your backend/ system with a frontend that surfaces all its capabilities
(screenshots of the review frontend of a tool I built in < 2 hours)
workflow that works well when vibe coding tools:
- build the backend with codex (workflows, backend, db etc..)
- /loop in claude code: "Every 30 min, rebase from main and rebuild the frontend to surface everything from the system"
workflow that works well when vibe coding tools:
- build the backend with codex (workflows, backend, db etc..)
- /loop in claude code: "Every 30 min, rebase from main and rebuild the frontend to surface everything from the system"
@PaulRBerg they beta tested this on me 2 weeks ago, was interesting (you can, with the arrows, get into them and message them which is really great) but it was messy as hell at the time