I manually approve everything (aside from the "allow commands like this" option). Not because I'm worried it's about to do something harmful but because I want to stop it from wasting tokens doing something stupid like implementing a feature I never asked ask for. I'm not even reading the commands, just checking it's going in vaguely the right direction.
I think my ideal setup would not involve approving commands, but approving summaries of proposed actions (whether they involve code exec or file edits). Kinda like plan mode but perpetual. Like most things, trying to solve it via CLAUDE.md is an exercise in futility.
I run in manual mode not out of rogue agent or token waste concern but because I want to effectively pair program with the agent. I need to understand what is happening along with the how and why. I’m confident I’m slower than YOLO people but that’s a trade off I’m consciously making so I can stay in the loop to comprehend the codebase.
Your summary approval idea is interesting and feels maybe like a mini plan mode. My biggest frustration with the existing manual approval system is when Claude is exploring it gets tedious to approve each command. Being able to approve a block of commands or a mini plan AND have auto mode audit them for safety would probably be something I would consider for the expiration phase of my Claude use.
I spent part of last week, and the weekend, in Claude (Fable) on a documentation project... maybe a waste to use Fable, but been a little obsessed with trying to get the technical details right. I know, from personal history, more than some of the available info easily found on the subject. I let a lot of it just go, and swing for the fences, but in the end, did hand review all the output at each stem.
In my specific case, implementing modern documentation for Remote Imaging Protocol (and old graphics protocol for terminal/bbs usage), and in particular trying to gather what exists of the later version released. The company itself folded, and there wasn't a lot of broad use of the later versions, and only partial information was ever released as a documentation. Claude was able to interrogate the apps, and some available open source, in progress work, and establish a lot of things.
I've still got to direct some edits... I guided a re-flow of the order the documentation was written in, particularly to be consistent between the versions... but it obsessively added verbose summaries to every page/file. All markdown, meant to build a website from, and eventually some compatibility tests, guidelines and a green/open implementation.
This morning I've turned auto mode back on, I'm only using it during Claude's exploration phases of my workflow and then I'm reading the full transcript (which includes some thinking info). So far, this seems to be quite close to what I want, and an improvement over my previous "manual all the time" mode.
Still leaving manual approval for all edits. Combined with reading the full transcript of the exploration, I feel I stay in the loop pretty well in this first test.
You can get some approximation of that with more detailed plans. I use a prompt like "Let's make a detailed plan. Write plan files as markdown to plans/feature-XYZ. [regular prompt]". That way you get much more thorough plans than with plan mode, and you can make it as detailed and deep as you want.
It can't account for things that only come up during implementation. But you can explore much more of the design space in planning to reduce the opportunities for Claude to wander down the wrong path during implementation
I've tried this, and it is largely what I do, but it's easy to end up with a large tangled pile of markdown files, and the implementation either diverges from what's in markdown or you burn a lot of tokens keeping them in-sync.
The plans usually end up very long and verbose, and in my most vibe-y test project I stopped even reading them myself and instead just ask questions interactively. e.g. "what are the remaining unanswered design decisions", "which part is jankiest". It does work quite well although I don't know how I feel about it.
P.S. to any anthropic staff: The "chat about this" option in Plan mode is useful but also annoying. When I select it I want to take the next turn of the conversation myself, I don't want it to waste time "thinking" just for it to verbosely ask me what I want to chat about.
I don't think there's any functional difference between "chat about this" and just directly entering what you'd like to do / ask about in the "Tell Claude what to do" option, I think that's pretty much exactly what you're looking for.
That's close to my workflow as well... keeping planned areas in TODO.md, done changes in DONE.md, and a non-committed WORKING.md for the current work being done, as well as coordination with sub-agents, so I can view/monitor... have to direct the latter each time to have it done that way.
Imho, the combination of the mattpocock skills, and the beads local issue tracker take care of all of that very well, with generally high quality output.
That's a good take. If LLMs are supposed to replace programmers, then they need to be programmers. Executing random things in all directions hoping for a break is something novices do. A good teammate knows when to ask, what to ask, how to ask, and who to ask.
> Not because I'm worried it's about to do something harmful but because I want to stop it from wasting tokens doing something stupid like implementing a feature I never asked ask for
I think my ideal setup would not involve approving commands, but approving summaries of proposed actions (whether they involve code exec or file edits). Kinda like plan mode but perpetual. Like most things, trying to solve it via CLAUDE.md is an exercise in futility.