Superpowers for DeepSeek Reasonix

One of my daily driver skills I use every day in Claude Code is arguably the most popular, Superpowers.

It’s subtly becoming the shape of mature agentic behavior within a codebase. I decided I wanted to port this over to DeepSeek’s new coding harness, Reasonix. The catch, though, is I wanted to see how well this suite of skills would elevate a flash-level model. Specifically, I built it to work well with deepseek-v4-flash.

The Main Skills:

• brainstorming: Before any build work: turn a rough idea into an approved design
• writing-plans: You have a spec/requirements for a multi-step task, before code
• test-driven-development: Implementing any feature or bugfix, before writing code
• systematic-debugging: Any bug, test failure, or unexpected behavior, before fixing
• verification-before-completion: Before claiming work is done/fixed/passing
• executing-plans: Execute a written plan inline in this session with checkpoints
• using-git-worktrees: Before feature work needing an isolated workspace
• finishing-a-development-branch: Work complete + tests pass → merge, PR, or clean up
• receiving-code-review: Acting on review feedback (from review or a human): verify before implementing
• writing-skills: Creating, editing, or testing Reasonix skills

Caveman Speak

I specifically wrote the skill instructions themselves, as well as the tiggers using caveman talk. My reasoning was two fold:

• I was betting the clear and direct language of caveman would activate better on flash model.
• I wanted to save tokens

Before I began building the skills, I built an activation benchmark of contrived scenarios to determine if the skills would activate and behave as I was expecting. Before caveman, my activation count was around 6/10. After re-writing all skills to caveman style, it was 10/10.

The skill is still a work-in-progress that I continue to dogfood myself.

https://github.com/christopherarter/reasonix-superpowers