Summary
Sharbel A. shares a hands-on walkthrough for building effective skill files in Hermes, an AI agent platform, drawing on a personal audit of 55 of the most popular publicly available skills on GitHub. His headline finding: 49% of those skills never specify when the agent should use them, making them effectively invisible to Hermes’s routing logic and causing them to sit idle while consuming token budget.
The tutorial breaks skill-building into five steps — picking the right job, extracting raw content from past corrections, writing trigger conditions into the description field, sizing the file for token efficiency, and deciding when to write code versus instructions. Sharbel reveals that his own primary skill file ran nearly 12 times larger than the median publicly available skill, consuming 6.19 million extra input tokens over its lifetime, and demonstrates how offloading rarely used sections into a reference folder the agent opens on demand is the single largest token-cost reduction available to most builders.
A key conceptual distinction in the video separates instructed steps from scripted ones: for any outcome with exactly one correct answer — such as arithmetic or token counting — natural-language instructions let the model improvise on every run, producing inconsistent results, while a hard-coded script is both cheaper and more reliable. By the end of the video, Sharbel provides the skill-auditing file built during the tutorial as a free download.
📺 Source: Sharbel A. · Published August 30, 2026
🏷️ Format: Hands On Build
