anthropics
`Yes, and...`, `Yes, but...`
Seeking the missing middle between these:
Would you like to proceed?
1. Yes
2. No, keep planning
Currently, it seems about half of my 2 responses begin with the words yes and or yes but. The semantics are jarring for me, and probably can't be helping the models either.
"I've sent a proposal the the user"...
"The proposal was rejected with message: 'do the proposal, with caveats'"
vs
"I've sent a proposal the the user"...
"The proposal was accepted with caveats:"
The plan mode "No, keep planning" option doesn't say "The proposal was rejected with...". It just allows you to provide more instruction for the model to create another plan. Are you seeing behavioral issues with the current options?
I guess this is an inevitable tension when supplementing the chat interface with prescribed inline menus. People are accustomed now to be able to bring their own tone and diction to interactions with models. Option 2 here is outside my regular habits of speech, so selecting it in a conversational flow feels like a new kind of existential jankiness.
The plan mode "No, keep planning" option doesn't say "The proposal was rejected with...".
You might know this, but I don't personally have knowledge of how these responses are routed to the models. EG, is it the result of a tool call? Does it get piped in as (part of) a user message?
Are you seeing behavioral issues with the current options?
Am I seeing Claude catastrophically misunderstanding my intent? No.
Am I confident that the little extra sand in the interactions doesn't degrade performance by a half percent? Also no.
I would also love to have a 'Yes, and...' option, not necessarily for planning mode, but for normal mode : Claude often suggests a change I find perfect, and it gives me a new idea. I often say No, just to write 'actually yes do this change, and also do this other thing i just thought of before continuing with the plan'
Or sometimes Claude suggest a very long change, that is almost perfect but just needs a minor tweek. Saying no means claude uses token for the change 2 times, whereas it would be better to accept the change and have claude re-edit the file it just edited. I often do the tweaks myself in another terminal, but then I can't really inform claude that I did these tweaks which means claude has to re-read the file...
Just to add another reason for a third option that is not "yes do it" nor "no, keep planning"... I sometimes want to say "Yes, but don't do anything. I'll take care of it. That's it for this plan". And I'd rather a "user rejected the plan" wasn't recorded. Basically, "You've done your job, pat on the head, close up shop. On the next thing..."
I would also dearly like this. I know saying "No, keep planning" kind of achieves what we want, but is seems this has been requested enough times in various duplicate issues (some closed off in a circular duplicate loop of each other!), that is seems that the demand is there.
For me, my most common one is:
"Yes, but add an explicit check in with me after each stage"
I don't need to see the plan again. I especially don't want to be shown a different plan that the last time so I have to read the whole thing through again. I just want to Claude to start working on the plan, but checking in with me after each stage. (I have also tried dozen of different ways to get Claude to always do this, but none have stuck reliably).
Another common one are things like:
"Yes, but stage 2 is unnecessary, and then stop before stage 4" "Yes, but I've already actually done stage 1, so start at stage 2" "Yes, and then also add unit tests at the end"
I don't need, or want, to see the full plan again, and doing so feels like a waste of time and credits.
Rather than know I will sometimes also say "Yes" and then hit escape to stop Claude in its tracks straight away. Or I say "No, keep planning" and then switch out of plan mode and say yes, but it seems Claude then still needs to use the plan exist tool before it can start working on the plan.
Both current options seem inelegant, but it is also the little bit of extra cognitive load it puts on us as a user every time we want to say "yes but/and", and instead have to think "hmm, can't do that, what option shall I use instead?" - ie: "Don't make me think!".
Personally I'd also be happy if there was just a user setting where we could set "start working on plans as soon as they are accepted" to false. As for me the times I want to add additional comments, but don't want to see the full plan again, greatly, massively, outnumber those where I want Claude to start on the plan straight away.
What also might help in this area (and is maybe a separate issue), would be if Clause wasn't so keep to use the PlanExit tool as soon as possible once in plan mode. It's so hard to encourage it to discuss and formulate the plan before jumping straight to presenting it to you.
