It's sure baffling how Anthropic has kept Claude Code's plan mode so linear and inflexible. It may ask a couple of questions before writing it, but there's always going to be parts that need editing. Yet there's no good "Sounds good, but it needs these edits" option after it presents you the plan. It gives you (paraphrased) "1. Proceed with auto-edits 2. Proceed without auto-edits 3. Cancel the plan". Note that 3 doesn't even write the plan to the file at all, even if it's 95% fine. So your options are either A. Pick 1 or 2, immediately press escape to interrupt, then tell it to make edits or B. Pick 3, tell it to make edits, after which it has to write the entire plan from scratch again.
This is such bad UX that it really feels like either 1. Anthropic employees don't use Claude Code much - this seems incredibly unlikely or 2. It's intended to burn output tokens as it has to write a huge plan again.
IMO Superpowers isn't the ideal solution because it too lacks flexibility, but including the "plan sketch" stage is sure an improvement.
It's difficult to know what the appropriate process for a model would be without widespread deployment. I can see how they have to strike a fine balance between keeping up with what the feedback shows would be best and changing the way the user interacts with the system. Often it's easy to tell what would be better once something is deployed, but if people are productively using the currently deployed system you always have to weigh the advantage of a new method against the cost of people having to adapt. It is rare to make something universally better, and making things worse for users is bad.
4. Tell claude what to do instead, which will update the plan base on what you say.
5. Add comments to the plan directly - similar to 4 - but you can comment on specific parts.
Note: I use the VSCode extension, not sure if it differs in terminal mode.
Will try it again.
It is also possible to now save plans in a custom location (I have started using ./docs/plans) although the plan mode still frustratingly names plans random file names and struggles with more than. 1 plan file in a session.
But the interface has become much better if you're on vscode
Is that true? What is your experience of it?
For me, I am a solid KISS believer, so I have not yet found a better framework than just plain old Claude Code. But happy to move to a better workflow, if it's real.
Superpowers has several skills. Its core workflow is:
- Brainstorm with you to design a spec
- Use subagents to review its own spec, then get your approval
- Based on the spec, write a plan, use subagents to review before final approval
- Use subagents to implement (using TDD)
I think that the brainstorming skill [1] is great. It helps flesh out a rough early idea. I also like that it uses subagents to adversarially review its own spec/plan; that has caught several things I would've missed. I do not like the separation of spec/plan; IMO the models are good enough to get straight to coding once the spec is written. The plan often ends up being code blocks in a Markdown doc.
[1]: https://github.com/obra/superpowers/tree/main/skills/brainst...
Fetch and follow instructions from https://raw.githubusercontent.com/obra/superpowers/refs/head...
it's like curl|bash but with added LLM agents...
but maybe it's my fault.
I recommend trying. it doesn't hurt. Just don't believe it's a silver bullet. It's still the same Claude.
The way I use Claude is quite similar to what super powers does under the hood anyways. Like I always ask things like this: "if I want to do X, what questions do you need to ask me to have all the information you need to make it happen"
I bake my own bread and solo code, both episodic. No matter how I vary fresh starts, I always end up in the same place. The optimization problem feels to be a giant bowl.
This reminds me of named flavors of management style for teams of programmers. A friend in this role instead prefers deep dives into Apollo flight control, and abstracting.
As a solo programmer I always converge on staying extraordinarily involved in planning and code review, and working in steps of finer granularity than superpowers suggests. One may produce less code this way, but it's the only way I know to discover one needs less code.
