martin.p @OnePromptMagic

@Berry_cooool auto

@cursor_ai benchmarks within cursor are not very interesting. a reql metric would be opus in claude code vs Composer in cursor. the agent orchestration around the actual model is the crucial part and it's not a fair comparison just within cursor

2 months in JP coming to an end soon. it's so weird how my timeline changed from entrepreneurship and builders to full on Japan twitter. I'm not saying I didn't enjoy the switch up once in a while, but not seeing everyone hustling kinda made me lose motivation to post a lot :D

@levelsio @arena if it wasn't hard before to tell what's real or not, it's not getting easier with this 🫠

learn from my mistakes: set reminders for when tokens or credits are about to expire. ur future self will thank you :D

a year ago I was actually commited af to n8n workflows, as they make entry and exit point to other applications really easy. nowadays I feel like it's not even worth opening the app if my agent can basically oneshot whatever I want 🙈 Shit's changing too fast

why is anthropic banning people's accounts for analysing the leaked source code? Seems like not the smartest move to make. Just accept you fucked up, and instead why not do a video on the architecture and how they built it with claude

it likely depends on your phrasing. if you simply ask for a review it likely 'wants' to find something (same as when I'm reviewing someones code I'm subconsciously trying to find an issue that might not be there). Try providing the plan, then saying 'here's my plan, are we good to go?'

@thekitze now reprompt in new session with "praise it properly but rost where it's due" and let's if that changes the rating

@blankoween @TukiFromKL yeah sure it's a good piece of software but people act as if they leaked their model weights or smth.

I'm ususally all about leting the AI do the work, the research, the thinking, while I'm doing the 'big' thinking. but lately Opus is just actually smarter than me in many if not most cases. That's why I sat down and read a 1000 pages of system architecture books. gotta keep up with the AI overlords, at least as long as I still can :D

coffee up, today's the day for building some wild shit ☕️

I just finished reading "Clean Architecture" by Robert C. Martin. In a future where agents take over all of programming, it will be crucial that clean patterns and architecture is in place. why? because agents will always look at how you've implemented your code so far and use similar style/patterns. E.g if you got loads of comments in your code the AI will continue adding comments. This will bloat the whole repository and make it less readable and maintainable. Same happens with bad code design, AI will rather multiply on your bad code than suggest a refactor, because it wasn't asked to. I then setup a claude skill that scans the repo based on my learnings in the book. Idea is to have a bigger refactor to adhere to best practises now, and later run the skill every few PRs, checking for clean architecture. I assume that after the refactor AI will however already do a cleaner approach than before due to the patterns it will find. Let's see how it goes

claude code has been amazing in helping me setup my sales pipeline, but in long sessions it often gets lazy and does things half assed. it would tell me "yeah I'm sure everything is ready to go" but when checking manually there still are many flaws or missing pieces. To solve this, I've implemented "santity-check" endpoints throughout the infrastructure that when called do all the checks I'd do manually. these provide a list of open issues I can then paste into claude to fix until all checks pass. the endpoint is alao available viam mcp so I setup a skill to do all sanity-checks and fix until green. this helped me a lot with concsistency in my AI outputs even for long sessions

I usually work with 4 claude code windows, and I'd try to not have them work on the same domain. like one builds on the product, one works on the pipeline, one is for trying things out etc. one thing that works for me is to change background color on terminals just so I can associate the context windows faster in my head. Another thing that works for me is to have all terminals on one window and NO browser open there. all of the tabs are open on another screen to prevent distraction while waiting for agents to finish. I'm definitely in the same boat as you and these are just some things that worked for me. Will continue improving on my flow and share it

as technical founders sales always comes short. we're not good at it (yet), but we're getting there. One hack we're applying right now for us to keep learning: one sales tutorial every 2 weeks with followup discussion + action items - works wonders.

@zuess05 same, I even got one of those old subscriptions where you'd still switch into slow mode instead of running out of credits. but it's just still so little usage on the god models (I'm out of opus 4.6 after a day max)

so many people complaining about claude limits being reduced. IMO this is clearly just people getting used to the new 1M context window. it's so convenient you just keep on writing in that one chat window. the underlying issue is tho, that with longer context come higher token costs, that's just how it is. restart your sessions from time to time and you're fine. I personally didn't see any decrease in limits at all.

@levelsio how are you doing the blocking process? I assume not manually? 🙈