Tabular Editor's guide to writing agent skills: encode the business, then cut
A Tabular Editor blog post explains how SKILL.md agent skills load, what belongs in them, and why to iterate by trimming rather than adding.
A Tabular Editor blog post published October 20, 2026 gives a practical walkthrough of writing agent skills. Its main advice is to put what the agent can't learn from training into a SKILL.md file, meaning your business's processes, definitions and preferences. Leave out the technical how-to the model already knows. It was shared on X by @kurtbuhler, who credits @SQLGene and @VdVoordeR as authors.
The post sets out how a skill loads. A skill is a folder with a Markdown file plus optional references, scripts and asset folders. A harness such as Claude Code or GitHub Copilot reads only each skill's name and description at session start. The full instructions load when you or the agent invoke it, and reference files load only when the instructions call for them. That makes the description the trigger. The authors warn that a vague one means the skill never gets used, while one that overpromises pulls it into the wrong tasks. They also note that pointing an agent at a skill file isn't the same as installing it, and that implementation "varies quite a bit by harness." For example, allowed-tools is still experimental.
Their suggested method is to have an agent attempt a task you already do and note where it guessed wrong. Write those notes up as a skill, then test again. Overly prescriptive skills can confuse the agent, so consider what to cut before adding more. The post includes a short bill-of-materials example skill. It says that measuring whether a skill actually helps is the subject of a part 2.