fainshare/liam-skills

Liam Skills is a set of thinking method skill packs meant to be loaded into AI agent platforms such as Claude Code, Qoder, or Aone Copilot. The main project inside it, called Taixu Wuzhuan (Five Turns of Taixu), turns a five step thinking process, diverge, converge, verify, review, and reclaim, into five separate skills that can be used one at a time or chained together. The problem it is trying to solve is that an AI agent tends to lock onto its first idea and reason around it from then on, and a single prompt cannot fix that on its own. Each of the five skills is named after a thinker and handles one step. Osborn spreads out as many possible ideas as it can, using structured prompts like substitute, combine, and reverse, so the agent does not settle too early. Hegel then runs a calibrate, challenge, and deepen loop over those ideas until they converge on a smaller set of real findings, including checking what evidence would exist if the original idea were wrong. Bayes turns each surviving finding into a checkable prediction and tests it against real code or data, treating any mismatch as useful information about a blind spot. Feynman reviews the resulting work for self deception using six principles, and does this with a separate agent that only sees the finished change, not the reasoning that produced it, so it cannot be talked into agreeing. Debono then looks back at whatever got rejected along the way and asks under what conditions it would actually have been valid, to catch good ideas that were thrown out too quickly. All five can be run together through an included pipeline tool with pause points between steps, or run selectively. A separate skill called Socrates applies a similar checking approach to outside claims, such as industry reports or competitor analysis, rather than the agent's own output. Suggested uses include evaluating technical proposals, reviewing architecture decisions, and checking investment or content arguments. It installs by cloning the repository and copying the skill folders into your platform's skills directory, and needs a Bash environment since some skills run shell commands to verify their findings. The project is released under the MIT license.