🤯 AI's Genius Hack: SkillOptis Unleashed! ✨

SKILLOPTIMIZATION: A FRAMEWORK FOR TRANSFERABLE LANGUAGE SKILLS
SkillOpt, a novel text-space optimizer developed by a team of researchers, offers a streamlined approach to training natural language skills. The core mechanism involves a single, frozen target model alongside an optimizer model that proposes bounded edits based on rollout scores. A held-out selection split determines edit acceptance based on score improvement. The final output is the `best_skill.md` file.

THE CORE MECHANICS OF SKILLOPT
The SkillOpt system operates through a distinct process. A single natural language skill document is trained, while the target model remains frozen. An optimizer model analyzes scored rollouts and suggests edits, constrained by predefined bounds. An acceptance criterion is defined: edits are only accepted if they strictly improve the score. The resulting artifact is the `best_skill.md` file, representing the optimized skill.

MODEL ARCHITECTURE AND TRAINING
SkillOpt leverages GPT-5.4 and GPT-5.5 model variants for training and deployment. Three rows are central to the research’s findings, each representing a different scale of model: GPT-5.4-mini, GPT-5.4-nano, and GPT-5.4. Training costs vary considerably, ranging from 0.6M to 46.4M training tokens per absolute test point, depending on the benchmark. The optimizer model operates solely during training, eliminating inference-time calls during deployment.

KEY OBSERVATIONS AND RESULTS
The research highlights several key observations regarding skill transfer. SpreadsheetBench on GPT-5.4-mini retains approximately 82% of its in-domain gain, representing a near-free reuse scenario. Conversely, LiveMath on GPT-5.4-nano exhibits a more unusual behavior, scoring 28.8 against an in-domain SkillOpt result of 27.2, suggesting that certain learned procedures are target-model agnostic. Retention rates are not uniform, with no row falling below the target’s no-skill baseline. The experiment focuses on four model families, with cross-family transfer (e.g., GPT toQwen) not tested. All rows utilize GPT-5.5.

TOOL AND API PORTABILITY
The Codex and Claude Code harnesses represent distinct tool and file APIs, exposing different command surfaces. A skill that survives this shift demonstrates a focus on structural and formula-aware verification, rather than encoding command recipes. This “structure-first inspection” approach is independent of the underlying Python CLI.

SKILL FORMAT AND TOKENIZATION
The final skills generated by SkillOpt range from 379 to 1,995 tokens across the six benchmarks, assembled from 1 to 4 accepted edits. The `best_skill.md` file supports this tokenization, and the paper’s Figure 4 samples one learned rule per benchmark, all of which are procedural.

TRAINING COST AND OPTIMIZATION STRATEGY
Training costs vary significantly, with SpreadsheetBench costing 0.6M per point and DocVQA reaching 46.4M. The optimizer model’s operation during training and zero inference-time calls during deployment contribute to cost efficiency. The Codex → Claude Code SpreadsheetBench result provides a compelling proof of concept, suggesting optimization should occur where tooling is cheapest and deployment aligns with product needs.

AUDITABILITY AND TRACKING
The deployed artifact is a text file that domain practitioners can readily understand, facilitating auditing. Every change is traceable via an `edit_apply_report.json` file, recording the status of each accepted or skipped edit. This approach contrasts with deploying fine-tuned weights, offering greater transparency and control.

RESOURCES AND REFERENCES
The research provides access to a comprehensive set of resources, including the original paper, GitHub repository, project page, documentation, PyPI package, and a demo video. Key baselines referenced include GEPA, TextGrad, EvoSkill, and Trace2Skill, alongside benchmarks such as SearchQA, SpreadsheetBench, DocVQA, LiveMath, and ALFWorld.