A Single Rewrite Suffices: Empirical Lessons from Production Skill Description Optimization

Yangqiaoyu Zhou, Mohammad Alqudah, Kwei-Herng Lai, Aaron Halfaker · 2026

Read A Single Rewrite Suffices: Empirical Lessons from Production Skill Description Optimization by Yangqiaoyu Zhou, Mohammad Alqudah, Kwei-Herng Lai, Aaron Halfaker (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-30775.txt.

Description

Yangqiaoyu Zhou, Mohammad Alqudah, Kwei-Herng Lai and Aaron Halfaker report lessons from optimising the descriptions that route questions to an enterprise AI agent's skills. When two skills have overlapping descriptions, the routing model sends questions to the wrong skill, a failure the team call skill collision. With nine skills and three hundred seventy-two regression cases on a production chat agent, they built a pipeline to rewrite descriptions.

The pipeline's descriptions matched hand-tuned ones, at about seventy-nine percent F1 either way. Engineering effort per skill fell from two hours to under four minutes. Ablations on that system and on the ToolBench collection isolated the drivers. A single rewrite by a language model, fed any false-positive and false-negative cases, captured most of the gain. Iteration budgets, feedback composition, paired editing and training set size moved the score by under half a percent.

Catalog record

CREATORYangqiaoyu Zhou, Mohammad Alqudah, Kwei-Herng Lai, Aaron Halfaker
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.30775

Browse the full library · About the Tower of Babel