About forty agent tools now read Anthropic's SKILL.md workflow format

AI Frontier Post's tutorial, republished on dev.to, gives a size for each loading stage ^[17]. Metadata runs about 100 tokens per skill and stays in context from startup ^[6]. The recommended ceiling for the instruction body is 5,000 tokens, and the body loads only when the model matches a task to the description ^[7]. "You can have fifty skills installed and barely feel it," the tutorial says ^[9]. At 100 tokens each, fifty skills put 5,000 tokens of standing context into every session, the same as one full-size skill body ^[10]. The 100-token figure is the tutorial's estimate. It holds only while each skill's metadata stays at the few lines of YAML the format expects ^[4].

The weak point is the trigger. Loading the body is the model's judgement call against the description ^[7]. When the model makes that call, the only parts of the skill in its context are the name and the description ^[6]. If the description is vague, a correct procedure stays unread on disk. I would spend more editing time on that one field than on the instructions below it. The frontmatter is what tells the agent what the skill does and when to use it ^[4].

The scripts/ folder is the best-designed part of the format ^[5]. A script's code never enters context. Only its output does ^[8]. A version lookup or a format check written as a script costs the tokens of its result. Written as prose in SKILL.md, the same step costs its full length every time the skill fires ^[7].

The portability figure comes from the tutorial: roughly forty products by mid-2026, with Claude Code as the reference implementation and a skill that "works in every agent that speaks the open standard" ^[3][14]. The required surface is one file with two mandatory fields ^[4]. Two fields are easy to agree on. For write-once to hold, each tool also has to stage loading the same way and run bundled scripts. The tutorial does not test that tool by tool. Installation also goes to a per-agent location, "the directory where your agent looks for personal skills," and for Claude that is ~/.claude/skills/ ^[15].

The tutorial sets three conditions for when a workflow deserves a skill: you have done it at least three times, it has multiple steps with fixed conventions, and the output has to come out the same way each time ^[11]. Single shell commands belong in aliases and pure reference facts in docs ^[12]. The tutorial also says each skill should do one job, and a skill that needs a table of contents is really two skills ^[13]. Building one takes about 20 to 30 minutes, with Python 3 and pyyaml for the validation script and no API keys ^[16].