wtf is skillopt-content?

skillopt-content is a small toolkit for improving the written instructions, or skills, that an AI writing agent follows, without ever touching the AI model itself. It applies an approach inspired by a research paper called SkillOpt, treating a skill document like a piece of text that can be gradually edited and improved through a repeatable, checked process, rather than something a person has to perfect by hand. The core idea works in a loop: the current skill document is run against a set of training articles, the system proposes a limited number of add, delete, or replace edits to the skill text, those edits are applied, and then the updated skill is checked against a separate held out set of articles. An edit is only kept if it strictly improves the score on that held out set, edits that fail are stored in a rejected edit buffer instead of being thrown away, so the system does not keep proposing the same failed idea again. The repository includes a generic edit planning skill template, a pre publish checklist for reviewing any piece of public writing, and a small before and after example showing what one accepted skill update looks like in practice. It deliberately does not include any organization's private voice rules, so those are meant to be kept in a separate private profile rather than committed to this shared repository. Importantly, the scoring system that ships in this kit by default is only a simple deterministic mock used to test that the loop's plumbing works correctly. It is explicitly not a real quality evaluator, so anyone who wants genuine improvement needs to plug in their own scorer, whether that is another AI model acting as a judge, human ratings, or real product metrics, before trusting the results. The project is released under the MIT license, a short permissive license that lets you use, copy, and modify the code freely as long as you keep the original copyright notice.