One Skill Takes Over: Clashing Agent Skills and a One-File Fix
A · Strong evidence USC researchers build a controlled benchmark with code-checked rules and reference outputs, test seven models from five families and isolate the cause and the fix, with public code and data.
Summary based on the full text·Original: about 45 minutes to read
In short
A preprint (arXiv, 18 October 2026) by Parsa Razmara, Seyedarmin Azizi and Ardalan Aryashad of the University of Southern California. Agent Skills are instruction files that tell a coding agent how to do one job. When two skills for the same job sit in the model's context, the model usually follows one in full and drops the other, including rules that conflict with nothing. The authors call this takeover. The worked examples inside the skill files drive it, and one merged file with one correct example fixes it on every model tested.
Key numbers
- 93 to 99%share of a skill's rules followed when that skill is loaded alone
- 61 to 62%share of a skill's rules followed when a second same-job skill is loaded beside it
- 1 to 15%takeover rate when each skill file ends with one worked example that follows both skills
- 96 to 99%achievable rules followed on all seven models with one merged skill file written by Claude Opus 5.5
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- SkillClash: 16 jobs (commit messages, PR descriptions, SQL queries, bug reports and others), each with two skills of six rules and one worked example. Code checks every rule.
- Rules followed: 93 to 99% with one skill alone, 61 to 62% with a same-job skill beside it. An unrelated second skill changes nothing.
- Most outputs sit in one of two corners: every exclusive rule of one skill met, none of the other's. This holds for a 4B model and for Qwen3-235B and GPT-4.1.
- When each skill file ends with an example that follows both skills, takeover falls to 1 to 15% and 90 to 96% of achievable rules are followed, without merging.
- Telling the model which skill has priority does not bring the dropped skill back on any model. "Apply both" works only on the two strong models and leaves interacting rules broken.
- A merged meta-skill written by Claude Opus 5.5 reaches 96 to 99% on all seven models. Most of its gain over a plain list of both skills' rules comes from its one worked example, which helps only when it is correct; removing it lowers the score by up to about 25 points depending on the reader model.
- A step-by-step merge depends on labelling rule pairs; Opus 5.5 labelled all 48 interacting pairs correctly, weaker mergers did not.
- Activation steering on three small models flips which skill wins in 38 to 49% of takeovers but makes the model follow both skills in at most 13% of outputs.
How they got it
The authors constructed the 16 skill pairs themselves so they could control how rules interact: one contradictory, one shared, one entangled and three exclusive rules per skill. Each pair runs twice, once with each skill as primary, against a reference output that meets all ten achievable rules. A co-author blind to the labels found all contradictory and shared pairs. Seven models (Qwen3-4B, Qwen2.5-7B, Mistral-7B, Llama-3.1-8B, Phi-4-mini, Qwen3-235B, GPT-4.1) were scored with 95% intervals over the 16 jobs.
Limits
- Skills are synthetic and short (six rules each); real skill files are longer and messier and come without rule checkers.
- The only real-world check is observational: in public run summaries from another study, the effect is about 8 points, smaller than in the benchmark.
- The best fix depends on one strong merger model; the authors flag that only one merger labelled rule pairs well enough.
Why it matters for companies
Any team that lets staff install skills or instruction files from several sources will hit this: a personal and a team skill for the same job, or a vendor and an in-house one. The failure is silent, since the output looks complete. The practical rules are cheap: keep one skill per job, merge overlapping skills once with the strongest model available, end the merged file with one correct example that satisfies both, and do not rely on a priority statement. Skill libraries need an owner who checks for same-job overlap.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org