A skill will work even with a first draft of the instruction — you can write in business or technical language, whichever is comfortable. But the quality and stability of the result depend directly on how the instruction is worded. This article collects the practices that make skills predictable.
The single most important rule. Compare:
"Open all pages that contain floor plans" — the agent executes this the same way every time.
"Open all pages where people live" — the agent may interpret this differently from run to run.
Call things what they're called in the product: floor plans, measurements, pages, groups. The agent shouldn't have to guess what you meant.
Don't pack several actions into one step. Three short steps are more reliable than one long one, and when something goes wrong, you'll know exactly where. See [Skill Steps: How to Describe Tasks].
Especially on the first runs:
Run the skill on a real page.
Check the output.
Find the step where the result drifts.
Tighten the wording of that specific step.
Don't rewrite the whole skill when one step misbehaves — fix it locally.
A skill isn't "write it and forget it" — it's a tool you fine-tune. The more precise the instruction, the more stable the workflow. A well-tuned skill performs predictably run after run, so a couple of refinement passes early on pay off in every future project.
Use If/Else before the steps where a mistake would cost the most — for example, before results feed into a report. See Using If/Else Branching for Quality Control.
