Comment by hadlock

I have a pretty specific workflow, I deal with live systems a lot (primarily in staging but occasionally prod) and have a really neat skill that has hooks into all the various things and then writes (&& checks in git) an investigation doc (.md) using a template into a git knowledgebase by month/quarter. That knowledgebase is really useful for frontloading the context of my new sessions as a lot of my work is iterative on previous work. It is really good at 80% of tasks, but simple stuff like using sops or modifying subcharts in helm it really struggles with. I do a lot of integration work between monorepos and Opus 4.7/4.8 really struggle with that kind of work. Opus is so bad at sops work I just outsource that to Codex/GPT5.5 now because the failure rate is so high.

The other high failure rate things is opus correctly predicting where endpoints will end up (stacked endpoints), when you have an ingress with a main monorepo but it's endpoints are the outputs (not just json regurgitated/passed forward).

I figured by having the clanker write it's own documentation about what it did, and then the next time saying "go look at invest doc 2026-06-29-TICKT_1234.md, do that again, but here's what we're doing differently and why..." and it would immediately grep the situation, but it doesn't. As alluded to earlier I've been exploring other models since the corporate blessed one is struggling to keep up.

Curious to see what you release.