xuansenpa1/skillrevise

SkillRevise is the code release accompanying a research paper about making AI agents better at reusable tasks. The core problem it addresses is this: when you ask an AI language model to write a "skill" (a reusable procedure for completing some category of task), the first draft is often imperfect. SkillRevise is a framework that watches the AI try to use that skill, identifies where it goes wrong, and then revises the skill based on what actually happened during execution rather than just asking the AI to try again from scratch. The revision process has three parts. First, a diagnosis step examines the execution trace (the record of what the AI did and where it failed) and produces a description of what went wrong. Second, a principle memory stores reusable lessons from past failures so the system does not have to rediscover the same fixes repeatedly. Third, a revision operator rewrites the skill with specific anchors tied to the execution evidence, producing a new candidate. That candidate is then actually run on the task again, and only kept if it performs better than what came before. This try-diagnose-revise-test loop can repeat up to a fixed number of rounds. The paper evaluates the system across three benchmarks covering planning tasks, software engineering tasks, and general skill-learning scenarios. Across all three, SkillRevise with three rounds of revision beats both running with no skill at all and running with a skill that was only written once without revision. The improvements hold when skills written for one AI model are transferred to a different model. The repository includes the Python package, a command-line interface, the benchmark data bundles used in the experiments, and documentation for running your own evaluations. You connect it to an AI provider (the examples use OpenAI) through environment variables, then point it at a task manifest file and specify how many revision rounds to allow. The primary language in the repository is PDDL, a planning description language used by some of the benchmarks. The framework itself is in Python.