Lisaskill

Creator guide

How to build skills people rent

A skill turns your professional judgment into files an AI assistant can follow. This page covers the format, what makes a skill genuinely good, and how publishing, versions and licensing work.

Anatomy of a skill

A skill is a folder of markdown files. The only required file is SKILL.md, which starts with YAML frontmatter:

  • name — the slug (lowercase letters, digits, hyphens).
  • description — the single most important line you will write. Third person, specific, and it must say when to use the skill, because the AI decides from this line alone whether to load your skill. Max 1024 characters.

Everything else is supporting material in references/*.md — checklists, benchmarks, templates — linked from SKILL.md.

---
name: severance-review-de
description: Reviews German severance agreements (Aufhebungsvertrag) for employees. Use when the user shares severance terms or asks what is negotiable before signing. Produces a risk report with payout benchmarks, deadlines, and negotiation points. Not a substitute for a licensed attorney.
---

# Severance Agreement Review (Germany)

## When to use this skill
- The user shares severance/termination agreement text.
- The user asks "is this severance offer fair?"

## Instructions
1. Identify: notice period, severance amount, garden leave, ...
2. Benchmark the payout using `references/benchmarks.md`.
3. Check every deadline in `references/deadlines.md` — missing
   the 3-week Kündigungsschutzklage window is irreversible.
4. Output: Summary -> Findings table -> Negotiation points -> Deadlines.
5. Always close with the "not legal advice" disclaimer.

## Out of scope
Executive contracts, collective agreements -> recommend a licensed attorney.

Progressive disclosure (why structure matters)

Assistants don't read your whole skill up front. They see the description, load SKILL.md when it matches the task, and only open the reference files SKILL.md points them to. That means:

  • Keep SKILL.md lean — under ~150 lines of precise, imperative instructions.
  • Move long material into references and link each file with a reason to open it.
  • Never bury the workflow: numbered steps beat prose paragraphs.

What separates a rentable skill from generic advice

A base model already knows generic best practices. People pay for your judgment. The skills that earn rentals encode:

  • Concrete checklists and decision criteria ("cap must be mutual, typical: 12 months of fees")
  • Red flags with severity ratings — things a generalist would miss
  • At least one worked example showing input → expected output
  • A specified output format (tables, sections, ratings) so results are consistent
  • Explicit scope: when to use it, what is out of scope, and when to escalate to a human professional (with disclaimers where your domain needs them)

Use the AI review in the editor's AI tab — it scores your draft against this rubric and proposes per-file improvements you can accept or dismiss.

Testing and fine-tuning your skill

You always test your current draft — publishing is not required. As the owner, every channel serves you your live working copy:

  • In the app: save a file, click Test in chat in the editor (or open Chat) and ask a realistic question. The assistant loads your draft SKILL.md and follows it — tweak, save, ask again.
  • In your own tools: connect Claude Code or OpenCode once (API key or OAuth). Your connection serves your draft, so every save in the studio is instantly live externally — no re-publish, no re-sync.
  • Renters are isolated from your experiments: they only ever get your last published version. Publish when the draft behaves the way you want.

Good test questions are the ones your future renters will ask — including messy, underspecified ones. If the assistant ignores the skill, your frontmatter description probably doesn't state when to use it clearly enough.

Publishing, versions and the lifecycle

  • Draft — your working copy. Only you see it; edit freely.
  • Publish — snapshots the current files as an immutable version (v1, v2, …). Renters always get your latest published version; publishing requires valid frontmatter.
  • Compare & restore — in the Publish tab you can diff any version against your draft and restore an old version into the draft, then publish it as a new version (nothing is republished silently).
  • Unpublish / archive — removes the skill from the marketplace and stops new rentals and renewals. People with an active rental keep access until it expires — they paid for the window.

How your content is protected

  • Paid skills are use-only. Renters can use them in chat and over MCP — where access and the rental window are checked on every single call — but they cannot download or export the files. Only you can.
  • Free skills download with a license. Files are watermarked with the licensee, license id and expiry, plus a LICENSE.md — traceable if they leak, and the notice explicitly tells AI assistants not to help redistribute the content.
  • Honest limits: anything an assistant can read, a determined user can copy piecemeal. The protection model is: no bulk export for paid content, per-call enforcement, watermarking, and clear license terms — not DRM.

Pricing and the AI builder

  • Free skills are great funnels — they get rented often and carry your name; keep your deepest expertise in paid skills.
  • Price in cents per rental period (you choose the period, e.g. 30 days). Renewals charge the same price again.
  • Starting from a blank page? Describe your expertise in the AI Builder tab — it drafts the full skill structure which you then correct with the things only you know. Then run the AI review before publishing.