Back to case study
SKILL.md validator

OpenClaw Skill Development

A practical study of OpenClaw's Skill system (teach an agent via SKILL.md, not code plugins), a complete Daily Briefing skill built from scratch, and a Lobster workflow chaining search → summarize → approve → push.

A live SKILL.md validator, not a replay. Paste or edit a SKILL.md; it parses the frontmatter in your browser and checks it against the Agent-Skills spec — name casing, the description token budget (≤250), metadata shape, trigger phrasing — and estimates the description's token cost.

OpenClawAgent SkillsSKILL.mdLobsterBash
OpenClaw Skill Development

Why this local version exists

The validator runs entirely client-side (parsing + checks + token estimate). The spec it enforces is drawn from the OpenClaw Agent-Skills course material.

Live · runs in your browser

SKILL.md validator — try it live

Paste or edit a SKILL.md. This parses the frontmatter and checks it against the OpenClaw Agent-Skills spec — name casing, the description token budget (≤250), metadata shape, trigger phrasing — and estimates the description's token cost. All client-side.

SKILL.md

Verdict

desc ≈ 51 tok

Valid ✓

Frontmatter

YAML frontmatter block found.

name

"daily-briefing" — valid kebab/snake case.

description length

~51 tokens (≤250 budget).

description triggers

Spells out when to invoke (good for routing).

metadata.openclaw

present.

requires.bins

declares deps: git.

body

Markdown body present (instructions for the agent).

The description is the highest-leverage field — it's what the agent matches against to decide whether to invoke the skill, and it lives in the system prompt, so its token cost is a real budget concern. Edit the sample above and watch the checks update.

What to try

Edit the sample SKILL.md and watch the checks update live as you type.

Delete the `name` field or break its casing — see it flip to a warning/error.

Paste a very long description and watch the ~token estimate cross the 250 budget.

What this demo proves

You understand the OpenClaw / Agent-Skills model: SKILL.md vs plugin vs channel, frontmatter routing, layered loading, token budget.

You can author a real skill end to end: tight triggers + a supporting Bash script + references.

You reach for structured orchestration (Lobster) with approval gates when free-form steps aren't reliable.

Model

SKILL.md (Markdown) teaches the agent — vs code plugins

Worked example

Daily Briefing skill: SKILL.md + collect-git-activity.sh

Orchestration

Lobster pipeline with a human-approval gate