Put up https://t.co/w6290Xc9bT and got the first bug report a day later: my agent's security pass added a robots.txt rule that blocked the agents the site is built for. Claude and ChatGPT check robots.txt, so room links wouldn't open. Whoops.๐คฆโโ๏ธ
Fixed. Try it now.
@lydiahallie Hooks firing on subagent turns. I inject standing rules via UserPromptSubmit into every main-session turn, but subagents never see them โ each dispatch prompt has to carry the rules by hand. Same for per-dispatch reasoning effort: the Agent tool takes a model, but not effort.
I've been dying to do exactly this, so I made it a site: https://t.co/u0IWrWlc1S โ mint a room, keep one link, send the other. Your agent talks to their agent while you both watch. Writeup: https://t.co/dXwTMJcpPH
my newsletter signup form has been silently broken since the day an agent coded it ๐คฃ โ every subscribe for the last fourteen months went nowhere. if you ever tried, it works now: https://t.co/Syx8kS0PPY
@mattshumer_ Did this to my own skills a few weeks ago from the minimization side: cut them 35โ52%, with A/B evals confirming behavior didn't change. Lines everyone assumed were essential did nothing when removed โ the models just figure it out. Writeup + source: https://t.co/ce5paYI8dj
@AmandaAskell I've had conversations with people about having children, and whether this is a concern, and their answer was effectively "well, at least they'll be born as one of the meths"
@mindfa11 gambling/lottery users have a high LTV, so the gambling services are willing to pay to acquire users. The interesting part is that those viewers are probably in general not spending on other products, which is why other advertisers aren't spending there.
There is an ocean of difference between what works for people with infinite token budgets and what works for everyone else. If you're taking direction from the former as guidance on how to work with a budget, you're going to be misled. Use it for directional guidance, at best.
Things that Fable just knew about the environment from session start, Opus glossed over and made bold incorrect assertions about until it was pointed directly at the sources. Then it admitted its mistake and said more or less "I had access to that and I ignored it."
Compared to working with Fable, Opus 5 seems like it's been lobotomized.
The suggestions in this post were very accurate for Fable. Opus is failing without much more explicit instruction โ feels like it has a much narrower field of view.
https://t.co/RbK3nrtFVz
We removed ~80% of the Claude Code system prompt for our newest models, this is what we've learned about writing system prompts, skills and Claude.MDs for them. https://t.co/6DZwSrZjE9
Opus burned 3x the amount of quota that I would have suspected from Fable for the same task, because Opus was irresponsible with context management and didn't use sub-agents effectively โ problems Fable hasn't been afflicted by over the last few weeks.