Discussion
I continue to use Claude Sonnet + Opus as the coding agents for my side projects but the inability to follow instructions is becoming a real pain.
I regularly have to deal with Claude deciding to ignore my instructions and make the app worse. Are there techniques or other models I should be using?
@carnage4life Got to use opus at work and i would prefer openai models for software engineering. They don’t get easily distracted.
@carnage4life watch context window and never let it get above 40%. I had Claude code customize its own ui to show me current usage per session.
@carnage4life while there is always dangers to anthropomorphizing, I do find treating Claude like a very bright but junior dev is effective.
Reduce the number of rules, and make them simpler.
Don’t rely on repetition or emphasis (besides typical MUST *when you mean it*). In short, don’t “berate” the model. Focus on being clear and consistent.
Move rule enforcement to a separate review stage rather than trying to do it all in a single context. One technique: Add checklists to commit template.
@cocoaphony @carnage4life what changes if you treat it like a capable but simple and limited machine?
Everything you say after still follows.
Thinking of it as a person is a step with implications.
@dimsumthinking @carnage4life I don’t find that predicts what it’ll do, though maybe there’s a specific kind of machine you have in mind. It definitely does not behave like my car.
I expect there’s a better mental model that’s more predictive, but I haven’t found it. I find “dehumanizing” it (avoiding polite language for example) is bad for my mental state. It might perform better if I were less polite, but I’d rather my habit of showing too much respect to objects than the opposite.
@cocoaphony @dimsumthinking @carnage4life I find myself using prompts like I would treat a colleague - and I love what you say about your mental state.
@carnage4life I find the agents stay on track better using the obra superpowers set of skills. You build a spec, then a detailed plan, then do TDD with sub agents. Opus 5.5 is the best model currently.
@carnage4life The trick I use is to just code things by myself without ever considering the hallucination machine.
@carnage4life I managed to put my Claude more on wheels, Opus 5.5 seems to be following instructions quite well for me these days.