A set of guidelines for designing agent skills—small instruction packages that teach coding agents how to handle tasks. It focuses on making their behavior predictable and allowing skills to work together.
★not rated 7 3mo agoA
tokens not measured
originalMIT
Design, build, validate, and publish production-ready Codex skills. Use when Codex needs to turn a repeated workflow or new capability into a new skill, organize its scripts/references/assets, forward-test it before first release, or package it as an installable repository.
Review, diagnose, and refine existing agent skills as living systems. Use when Codex needs to explain why an existing skill misfires, assess its architecture or evolution, apply an explicitly requested minimal improvement, recheck prior findings, or audit drift across source, published, and installed copies.
Use when a user wants to design one bounded feedback loop from a goal, assess or upgrade an existing Agent Skill with feedback loops, distill a completed conversation or work record into a candidate workflow and, when compatible, a loop-aware Skill, build an accepted Loop Craft JSON definition, or check an existing…
Use when reviewing scientific and technical manuscripts written in LaTeX; checking LaTeX source, equations, derivations, notation, appendix formulas, cross-references, citation keys, compileability, implementation ambiguity, unsupported claims, writing issues, or generating a LaTeX review file.
Use when reviewing, proofreading, auditing, or authorially polishing scientific and technical manuscripts in Markdown, DOCX, PDF, LaTeX, or plain text; checking research logic, methods, evidence, citations, reference relevance or duplication, AI-trace candidates, text clarity, terminology, style consistency, or…