Hi all,
I've decided to post this because I saw a post earlier telling people to engineer their own prompts rather than pulling something off of GitHub. I disagree. I think that pulling in the right skill at the right time is prompt engineering. Furthermore, us open source engineers work hard to give you our best. And following that line of thought, I decided to showcase my most useful skill: steelman.
A pattern I kept hitting working with Claude:
- Claude can be lazy. Once it's leaning toward a stack or pattern, subsequent reasoning builds on that instead of interrogating it.
- I was burning real time outside the loop, doing manual "is this actually right?" research before committing to whatever direction Claude proposed.
So I built /steelman. It takes the direction on the table (whether Claude proposed it or I did) and argues, at full strength, for the 2-3 strongest alternatives, grounded in project context. Then it makes an honest call about whether the leading choice held up.
What I noticed:
- Claude is simply better. Once alternatives have been weighed, Claude's downstream suggestions stop anchoring on the first idea. Sometimes it'll say something along the lines of "there was no reason to go with the first option except for wanting to avoid implementation."
- Working with Claude is faster. The "wait, should I double-check this?" moment disappears. The check is one command instead of an hour of side research.
- I find it extremely useful when using superpowers or brainstorm and Claude pitches 2-3 ideas. The steelman then pits them against each other.
I built it with TDD against 5 baseline scenarios. In the spirit of open source, it's MIT license and I would appreciate collaborating on it.
If this does well, I'll add it to the skill marketplace! And I'll share more. Looking forward to your using it and your feedback. I hope you genuinely like it!