A production agent harness must discover and rank the most suitable skill from a growing library for a user's task. At small scale this selection happens in‑context: the LLM planner picks among skill representations exposed in its system prompt without an explicit embedding‑based retrieval step. We treat this in‑context selection as the small‑N counterpart to large‑scale embedding retrieval and examine how Tinycloud, a multimodal video agent harness, represents its skills for the planner.
The harness ships skills in two recurring forms:
- Tool‑skills wrap a single external API or system tool and serve as primitive vocabulary;
- Workflow‑skills orchestrate multiple tool‑skill calls and render a template to produce a named deliverable.
These skills are exposed via two surfaces in the system prompt: an inlined‑body surface that provides full instructions, scripts, or templates for autoloaded skills, and a one‑line listing for on‑demand skills.
A six‑task selection ablation was run across three exposure regimes—all‑on, default, and all‑off. The findings are:
- With all‑on, the planner selects the gold skill on every task;
- All‑off slows execution and leads to hard discovery failures;
- The production default misroutes one task because the lexical signal of an autoloaded tool‑skill collides with a listed workflow‑skill, pulling the planner’s attention away.
The headline insight is that in‑prompt exposure of skills is not monotonically beneficial: partial exposure can create lexical competition that suppresses correct selection. We connect this small‑N observation to recent retrieval‑based skill‑routing work at large scale, framing the contribution as a case study rather than a benchmark.
Blogger's Review: This study highlights the subtle trade‑offs in prompt design for skill‑driven agents. Over‑exposing skills can backfire due to lexical interference, so practitioners should balance autoloaded and on‑demand listings, especially in multimodal contexts.