> Opus will fake out, say everything is done, and then you see that half of the plan was deferred, half of the functions are ridiculous stubs, ...
replyDoesn't Claude Code have a /loop command? Give it a message to keep it on track overnight, send every 20m, make it track progress in a doc, reread the doc after every loop. I've found this works well for a certain class of problems, most importantly where the actual work is getting done by very narrowly focused batches of subagents, with the main session just coordinating and keeping the doc updated.
They added a "/goal" command which I guess spawns a supervisor agent process that checks to see if your goal statement has been achieved (e.g. "/goal complete tasks 1-250 of plan.md") I've been pretty happy with it but I rarely use that workflow. Most of the time I give it a 3-6 step prompt and come back in 20 min and the first two were done and I get a summary "up next is to complete the next steps" which.... Opus 4.6 didn't have this problem. 4.8 feels like a cost cutting measure, or maybe it's just tuned poorly for my specific workflow (multi-repo system integration)
replyFor optimizations or proofs I suppose? Wouldn't know why else you would do something like that.
replyI think the parent comment stands - I’ve asked Opus to do a review of DeepSeek’s test suite and told it a couple things I wanted it to look for, and it did a very thorough review of the tests and picked out a reasonable number of gaps and tautological tests. It’s a mix of prompting/instructions, the agent harness, and random chance. The model is not wholly irrelevant but IMO increasingly so.
reply