what is pangu fr?

pangu-skill is a tool for evaluating and improving Agent Skills, the instruction files that give AI coding agents like Claude Code a repeatable procedure to follow. It treats a SKILL.md file as something worth continuously improving, not as a document you write once and forget. It sets up test prompts and a baseline score, makes small targeted edits to fix weak spots, checks the results through real or dry runs, and then asks the user to confirm whether to keep the change or roll it back. The project is not a simple formatting checker. It cares about two questions: is the skill written clearly, specifically, and in a way an agent can actually act on, and does the agent's real output actually get better once the skill is applied? The author points out that as teams accumulate more and more skills, manual review tends to miss problems such as documentation that looks complete but skips over failure cases, instructions that are long but lack concrete parameters and checkpoints, or a skill that scores well on format but does not improve real task output compared to a baseline. The evaluation itself uses a nine part scoring system covering things like the clarity of the frontmatter, how clear and executable the workflow steps are, whether failure modes and recovery paths are written down, whether there are explicit checkpoints before key decisions, and whether references and scripts are reachable. The heaviest weighted category is actual measured performance improvement, since a nicely written skill that does not change real output is treated as not truly effective. The whole process runs in five phases, from setting the scope and writing test prompts, through baseline scoring, to a validation gated editing loop where only improvements that pass verification are kept, using git as a safety net to revert regressions, and ending in a summary report of what changed. The repository includes the skill definition itself along with test prompts, reference material, and result card templates, plus a screenshot script built on Playwright. It is installed with a single command using the skills package manager, and it is released under the MIT license.