@aarondfrancis Love this - am working in a similar way with Claude Code but not as slick. The bit I still struggle with on big stuff is breaking up the review into sensible chunks. Branch / PR stacking just doesn’t work well with me. Do you do anything to make the big review more manageable?
@OBayesShelton@heytjmiller Anthropic rate limits for output tokens are much lower, so if you input and output is similarly sized you will probably just need a rate limiter for output.
@DaraghOShea@heytjmiller Do you need raw request and response (versus prism requests and steps) for prod, or just dev?
@heytjmiller were throwing this around the other week, and how we make this better without adding too much overhead.
@christophrumpel We’ve also just tagged a release with search grounding, which is very cool and worth a play if you are looking for more general info/research rather than specific data points.
https://t.co/d1a09Szooi
@christophrumpel Your context will get big quickly though as previous responses are included in message history with tools.
The other way to do it would be without tools. Send over the first pages html, use structured outputs to extract all links.
Then iterate over them with a second prompt.
@fabiomlferreira Two separate APIs for us. We’d do it that way even without Inertia as the data we want to expose etc. is different.
Anything shared can be put in services, actions, etc.
@Philo01 We do yea. I have checked our configured maintenance window and it is indeed at 3am Saturday.
Their docs say there shouldn’t be downtime, just increase in latency. So it is possible if we pushed timeout it’d be fine.
We’re dropping 1 transaction a week tops, so not too bothered