Meeseeks — Agent-Agnostic Skill Workbench

What this is

Meeseeks turns repeatable computer workflows into structured skills any agent can run. The same JSON drops into Claude (Agent Skills format), GPT (system prompt + tool calls), Gemini, or a local model — no translation layer, no per-agent rewrite.

The pipeline

Author a skill in a small Zod-validated JSON shape with bbox / accessibility-tree / success-check grounding.
Bundle it: derive SKILL.md (Anthropic Agent Skills frontmatter), manifest.json (grounding metadata), and actions.toon (token-oriented action list).
Run through the Playwright executor — accessibility-first resolution, screenshots per step, JSON receipt.

Why it matters

Today, every agent platform reinvents skill capture: ChatGPT custom GPTs, Claude Skills, Operator macros — all incompatible. Meeseeks is the smallest spec that keeps a skill portable while preserving the grounding metadata that makes execution safe.

Try it

Click ✏️ Author to paste a skill JSON (or pick a sample), see the plan with placeholders substituted, then 📦 Bundle to see the three derived artifacts. 🧾 Receipts shows real screenshots from the executor walking through these skills on a mock paralegal app.

Skill JSON — paste or pick a sample

Plan output

Plan results appear here after you click Generate plan.

Pick a sample and click Export bundle.

Real screenshots from the Playwright executor running these skills against the local mock paralegal app. Each pair is the page state before and after the step's action. Resolved via accessibility role + name (the gold-standard grounding path) — no bbox or text fallback needed.