What if the implementation is not satisfied?

[caibirdme]

Should I directly tell the agent, or should I use /specify.xx + problem

[wilsonneto-swe]

Sorry for the question, but could you please elaborate more on "Implementation not satisfied"? Did it implement but not everything? Didn't it generate the implementation or tasks steps?

[caibirdme]

implementation is the generated code from AI. Even though I use specify + plan + tasks + analyze + implement, the final output is not good. And then what should I do? I just give the feedback to claude code, but it seems that this feedback information wouldn't reflect on spec/files, which means if I compact the conversation or I change to another AI, the feedback will lose

[ebramirez]

One thing we have been experimenting in our team is to tell the AI (Claude), after a session of back and forth, to use the information of the session to improve the rules files that may apply.

[anthonyaquino83]

I have the same doubt, because yesterday I specified a CRUD app but in the end the edit task was not done, there was a //TODO in the code.
I just type in roo code "you didn't implement the edit feature" or do I go back to tome speckify.plan, speckify.implement, etc?
And if I want to refactor something in the code like:
Do not use *ngIf, replace all occurrences with @if?

[wilsonneto-swe]

It can be the case for reseting the context by opening another chat screen with fresh context and ask again /spekify.implement continue the implementation please.

Also, you can and should also edit the files, if you need update the tasks file manually highlighting that some points are missing and also sometimes, in these cases of context rot, we even need to manually update the concluded tasks in the tasks.