Many of us here are building and using Claude/Agent skills. I went looking for a set of thinking skills for a client and couldn’t find one, so I took a stab at building them. As some of you might know, I tend to overthink everything I do. This exercise was no different.
The problem they address: I’d ask a model something, get back a confident, well-organised answer, and stop checking. Microsoft and CMU found (The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers) the same pattern at scale: the more people trusted the output, the less critical thinking effort they applied. A separate paper found AI advice makes us less willing to say “I don’t know”, even when the advice is wrong and accuracy is rewarded. The better the tools get at sounding right, the less we check.
So these three push the other way. They ask one question at a time and use your answer to pick the next one. No summary and no tidy conclusion. You do the work, it guides you through the process.
- Systems Thinking For problems that keep coming back. Traces causes upstream, finds the feedback loops holding the situation in place, draws a causal loop diagram when it helps
- Critical Thinking For a claim or decision you’re about to act on. Paul and Elder’s Elements of Reasoning and Intellectual Standards, asked as questions, starting from wherever your thinking already is
- Critical Thinking for GenAI For text a model produced. You can’t walk the author back through the reasoning because there was no author, so this one runs eleven probes against the output itself
I have a background in using Critical Thinking, Systems Thinking etc and even I get benefit from these skills. A small example. I ran /critical-thinking on whether to buy a $250 online course. Four questions in, it had me compare it against the a course that stuck for me (one of Mike’s). That one had small-group accountability, a relationship with the facilitator, and a real external need. The $250 course one had a cohort of 1,000 and a pricing deadline that expires the moment you pay. I passed. Left to my own devices, I might have bought it.
npx skills add mlevison/agent-thinking-skills
Works with Claude Code, Codex, Cursor, OpenCode and around thirty others, and there’s a Claude Code plugin route as well: github.com/mlevison/agent-thinking-skills. CC BY-SA, so adapt away.
The overthinking shows up in places like this: the skills never let the model speak as “I”. It asks you to “provide an example”, never “give me an example”. A model that says “I think” invites you to treat it as a colleague who has weighed the evidence, and that’s exactly the moment you stop checking its work.
If it’s useful, star the repo so other people can find it. And tell me where it falls over.
I expect to expand the set of skills over time, to include cognitive biases and mental models - in so far as they get in the way of thinking clearly.
Help me figure out what to add next.