Write clear goals / map. Create ways to know when goals have been achieved. Create a way for agents to surface questions to you. Have a process that learns how you answer and start fast-laning decisions.
prompts like:
'Verification and tests should check that things work the way they are supposed to, not how the code under test is written. If it's unclear how things should work, consider it. If you are below 80% certainty, ask. Otherwise, plow forward.'
@sadkatwt I think using something like convex or supabase auth for less than $10 / month makes way more sense than rolling your own. I don't want to know about auth I want to build agentic flows that help people.
@_lopopolo You have carbon and astra ultra prompt your fable? I prompt my fable 5.2 to drive the rest. I think carbon is great with multimodal. Astra / Sol / Carbon for workhorse (depending on validation amount).
If you have prs piling up you must build tooling to lower the burden. Create ways to verify as much as possible. Separate prs into trivial, within expected abstraction boundary so light check, or big change need to have an expert understand. Otherwise you are in the way.
@Yuchenj_UW A quote or adage can impact our own minds and thought processes. I think agents (although much different from minds) are the same. The right 10-100 words at the right time significantly impacts long running workflows.
Opus 5.5 is better than me at coding and code review. It's better at summarizing. It's bad at choosing goals for soft-verifiable tasks.
The most constructive thing you can do is set goals and explain what you are looking for.
Harness etc can just be 3-5 good sentences
I ate one of my favorite meals today while listening on the phone. The food literally tasted less or lacking I think because listening took so much brain. It's like when you turn down the music when the driving gets intense
Opus 5.5 in vanilla Claude Code is so good that prompt engineering is back. The right sentence is like a spell you can repeatedly chant to make your software work.
@Vtrivedy10 I'm trying to build something similar for eng teams using multiplayer agent systems. What sort of insights / extractions provide the clearest value?
@ankrgyl We are trying out classifiers for sessions, hard coded rules, only access to processed artifacts, and RBAC for teams. How does the agent only access help?
Separately do you store annotations separately or in the same lake?