Installing useful instructions separately does not guarantee they will work together. Parsa Razmara and colleagues test what happens after a model has already loaded two skills for the same job.
Their SkillClash study constructs 16 pairs of instruction files, covering jobs such as commit messages, bug reports and SQL queries. Each file contains six rules and a worked example. Code checks every rule, and reference outputs demonstrate how compatible requirements can be combined.
- Across seven models from five families, per-skill rule compliance falls from 93–99% alone to 61–62% with another same-job skill.
- In 45–78% of outputs, one skill takes over completely: the model follows its independent requirements and drops the other’s compatible ones.
- Simply declaring a priority leaves much of that loss intact. Putting both original examples into one file also preserves substantial interference.
- Giving each file a correct example satisfying both skills reduces takeover to 1–15% across six tested models.
- A merged file written by Claude Opus 5.5 reaches 96–99% of achievable rules across all seven readers. Entirely correct outputs are lower, at 68–94%.
The example carries much of the repair. Correct examples from other mergers help similarly; incorrect examples add little or nothing. A structured merge can make matters worse when the merger mislabels compatible rules as contradictions.
The public code and recorded outputs make this a concrete regression test for instruction libraries. The practical experiment is to check one combined example against both source files before loading it.
Caveat: These are author-reported results on constructed text tasks, so the percentages do not measure reliability on real repository work.