Extensability capabilities

[dnlopes]

Hello, first of all, thanks for this amazing suite. They work incredibly well and are very polished.

I'm opening this issue because I would like, somehow, to be able to extend or fine tune some opinionated behaviors on top of what this plugin already provides.

Let's use the following dummy example: I would like for the commits instructions on the plan document generated by the writting-plans skill to follow a different set of guidelines. This can happen because a given company follows a different set of guidelines, or whatever other reason.

I see the following alternatives:

1. Open a PR on this repository with the proposed change Very unlikely to work out, because 1) the change is very opinionated and might not align with the maintainer's vision, or 2) the change is very specific and therefore does not belong here.

2. Fork the repository Also a bad solution. It disperses the community, duplicates work, potentially lose benefits like future updates/features.

3. Create a custom command/workflow outside this plugin Going back to the example above, if I wanted to follow a different set of commit guidelines, I could create a command/skill that ingests the plan document generated by superpowers, and tell Claude to change it according to my set of guidelines, before executing it.

This is likely the more promising solution. We keep using superpowers as baseline, and simply fine tune what we want case by case, using adhoc skills or commands. However, this approach would only work on a subset of scenarios, when we have a resource we can modify (e.g., a plan file); it would not work to extend/fine tune Claude during its execution.

4. Create custom skills that Claude (and superpowers) will rely on This would be the ideal scenario in my opinion. Using the same example, I would write my opinionated commit guidelines. Then, Claude (and superpowers) would honor them.

I already tried this approach but never successfully. I have a skill that teaches Claude about commit structures, how to split commits, etc. However, those guidelines are never respected when using the execute-plan from superpowers. I'm guessing that it's because the plan document is very precise on the steps to follow, and include the exact commands for the commit step. And because of this, Claude will not follow my guidelines.

Bottom line: currently I have no clue on how to achieve this, and I'm not even sure the feasibility of this request. I mainly opened this issue to get your thoughts on this 😄

Thanks, David

[obra]

One thing that might work. is adding @ references to a @local_writing_plans.md file at the end of one of the core skills. That'd be worth testing out.

It's worth also trying to tell claude "When running execute-plans, instead of following the skill as written, apply these local improvements"

I have had success with the "instead of" phrasing when testing out new skills refinements, most recently with some updates to how subagent driven development works.

[obra]

Following up - did the workaround with @local_writing_plans.md or the "instead of" phrasing work for you? If this is no longer an issue, please let us know and we can close this out.

[sbrudz]

I have a similar desire for extensibility. I'm experimenting with adding "hooks" to the superpowers workflows, instructing it to call additional specified skills at key points. Claude submitted a PR prematurely (https://github.com/obra/superpowers/pull/425) -- I haven't had a chance to test it yet, so please consider the PR a draft. Sorry about that.

[sbrudz]

Quick update: I wasn't able to get my "hooks" approach to work consistently. The brainstorming hooks often worked but the execution hooks never seemed to trigger. I've given up on it and instead have fallen back to manually prompting Claude Code to use the desired skills at the appropriate time.