Interesting how, when I use opencode + glm-5 to give feedback on things designed in multi-step by Claude Code + Opus 4.6, with a baseline prompt template and a single skill to connect to the forge, it finds shittons of relevant things that I didn't get from gpt-5.4-high or opus itself, unless I engage in endless prompt tuning.
Prompt template:
prompt = inline`/forgejo (context: ${org}/${repo})
Review the proposed code structure in issue ${ref}, issuecomment ${commentref}
from a robustness perspective and provide findings in a new comment to that issue.`
Relevant findings (on just a bunch of classes with only constructors and stubbed methods): 8/9. Only 1 finding was a misinterpretation of the desired functionality.
And the cherry on top: it gave Claude a bunch of compliments for great design decisions, lmao.
Interesting how, when I use opencode +
glm-5to give feedback on things designed in multi-step by Claude Code + Opus 4.6, with a baseline prompt template and a single skill to connect to the forge, it finds shittons of relevant things that I didn't get fromgpt-5.4-highor opus itself, unless I engage in endless prompt tuning.Prompt template:
prompt = inline`/forgejo (context: ${org}/${repo}) Review the proposed code structure in issue ${ref}, issuecomment ${commentref} from a robustness perspective and provide findings in a new comment to that issue.`Relevant findings (on just a bunch of classes with only constructors and stubbed methods): 8/9. Only 1 finding was a misinterpretation of the desired functionality.
And the cherry on top: it gave Claude a bunch of compliments for great design decisions, lmao.