A team spanning Microsoft and three Chinese universities has cracked a problem that’s been quietly plaguing AI deployments: skills that work in one coding harness often fall apart in another. Their solution, SkillOpt, doesn’t touch model weights. Instead, it trains a single markdown file.
The process is almost meta. An optimizer model reads scored outputs and proposes small edits—add a line, delete a procedure, replace a rule. A held-out selection split only accepts changes that strictly improve performance. The final artifact, a file called best_skill.md, typically runs between 379 and 1,995 tokens and takes just 1 to 4 accepted edits to build. Training costs are real but one-time, ranging from 0.6 million tokens per test point on SpreadsheetBench to 46.4 million on DocVQA. The optimizer model adds zero inference overhead at deployment.
The headline result lands in the cross-harness transfer. A skill optimized entirely within Codex lifted Claude Code’s score on SpreadsheetBench from 22.1 to 81.8. That’s slightly above the 80.4 Claude Code achieved by training its own skill from scratch. The two environments expose different tool APIs, different file surfaces, and different command structures. What transferred wasn’t command-line recipes. The paper points to workbook-level procedures—structure-first inspection, formula-aware verification, static-value materialization—that don’t care which CLI runs the Python.
But the portability has clear boundaries. On LiveMath, a reasoning-heavy benchmark, the same Codex-to-Claude Code transfer retained only 10% of the in-domain gain. And all cross-model experiments stayed within the GPT-5 family. Transfer to Qwen or other model families wasn’t tested, so whether procedural skills survive a shift in the underlying model’s reasoning style remains an open question. The paper’s bound is narrow but honest: no transferred skill ever fell below the target model’s no-skill baseline. The real story isn’t universal portability. It’s that a readable, auditable text file—one a domain expert can review in minutes—can sometimes travel further than a fine-tuned checkpoint.
