What Are Portable AI Skills? The SKILL.md Format Explained
A portable AI skill is a single markdown file that packages everything an AI agent needs to execute one business function — the frameworks, templates, tactics, quality checklist, and the rules for when to hand off to a human. Load the file, and any agent becomes an instant practitioner of that skill. No proprietary format. No platform to rent.
The word doing the heavy lifting is portable. Most "AI for business" products bake the intelligence into their platform: the prompts live in their database, the workflows run on their servers, and the day you stop paying, your operational knowledge evaporates. A skill file inverts that. The knowledge is a text file you own. The agent — Claude, GPT, a custom bot, whatever ships next year — is just the executor.
What's actually inside a skill file?
Every skill in the OSLO library follows the same structure: 19 required sections, in the same order, every time. That consistency is what makes 285 skills usable by an agent instead of being 285 blog posts:
| Sections | What they give the agent |
|---|---|
| Overview, Quick Start | What this skill does and how to start executing in minutes |
| Core Frameworks, Implementation Tactics, Templates | The actual how — models to apply, moves to make, artifacts to produce |
| Voice Commands | Natural-language invocation — say the skill's name and the agent knows what to deploy |
| Quality Checklist, KPIs & Metrics, ROI Calculator | How to know the output is good and whether it's paying for itself |
| Integration Points, Related Skills, Automation Hooks | How this skill connects to the rest of the system |
| Escalation Criteria, Human Handoff Points | When the agent stops and a human decides |
| Common Mistakes, Advanced Techniques, Industry Variations, Tools & Resources, Version History | Depth: the failure modes, the edges, the changelog |
Why does the same structure matter so much?
Because agents parse structure, not vibes. When every skill has an Escalation Criteria section in the same place, an agent running any of the 285 skills knows exactly where to look before making a call it shouldn't make alone. When every skill has a Quality Checklist, output review becomes mechanical instead of subjective. The format is boring on purpose — boring formats are executable.
It also changes what "training an agent" means for you as the architect. You don't fine-tune a model or build a custom GPT for every function in your business. You hand the agent a file. Swap the file, swap the expertise. Version the file in Git, and your business's operational knowledge has a commit history.
What does "agent-agnostic" buy a founder?
Three things that matter at the $5–50M scale:
- You're not betting the business on one vendor. The model landscape reshuffles constantly. Skills written as markdown ride every wave; skills locked in a platform sink with it.
- Your team and your agents read the same document. A skill file works as an SOP for a human and an execution spec for an agent. One source of truth, two kinds of workers.
- Compounding instead of restarting. Every improvement you make to a skill — a sharper template, a new failure mode in Common Mistakes — persists across every future agent that loads it. That's leverage that survives tool churn.
The skill is the knowledge; the agent is the executor. Separate the two and you own the part that compounds.
How do skills map to OSLO?
Every one of the 285 skills lives under one of the four OSLO domains: roughly 40 under Offers (offer architecture, pricing strategy, value stack design), 45 under Sales (scripts, objection handling, high-ticket closing), 100 under Leads (SEO content, paid media, email sequences, lead magnets), and 100 under Operations (SOPs, onboarding, dashboards, CRM, compliance automation).
That mapping is the point. A skill library without a sequence is a junk drawer — you'd still face the "which of these 285 things do I do first?" problem. OSLO supplies the order; the skills supply the execution. When the diagnostic says your bottleneck is Sales, you don't brainstorm — you pull the Sales skills and put agents on them.
The skills concept comes from FAST — the Factory of Agents with Skills and Tools. OSLO is literally the Skills column of that factory: the taxonomy that guarantees no business function is left without an executable skill.
What a skill is not
- Not a prompt. A prompt is a one-shot instruction. A skill carries frameworks, templates, metrics, and escalation rules — the difference between telling someone "write a sales email" and handing them the playbook.
- Not a course. Courses teach you so you can do the work. Skills equip agents so the work gets done while you architect.
- Not automation software. There's nothing to install. If your agent can read a text file, it can run a skill today.
FAQ
What is a portable AI skill?
A portable AI skill is a single markdown file (SKILL.md) containing the frameworks, templates, tactics, quality checklists, and escalation rules for one business function. Any AI agent that reads text can load it and execute that function immediately — no proprietary format, no vendor lock-in.
Why are skills written in markdown instead of built into a platform?
Because markdown is portable. You can copy a skill, fork it, version it in Git, and run it on Claude, GPT, a custom bot, or whatever agent comes next. Skills baked into a SaaS platform die with your subscription; markdown files are yours forever.
What are the 19 sections in a skill file?
Overview, Quick Start, Core Frameworks, Implementation Tactics, Voice Commands, Templates, Quality Checklist, Integration Points, Escalation Criteria, KPIs & Metrics, Tools & Resources, Common Mistakes, Advanced Techniques, Industry Variations, Automation Hooks, Human Handoff Points, ROI Calculator, Version History, and Related Skills.
How do skills relate to the OSLO framework?
OSLO is the taxonomy and the sequence; skills are the units. All 285 skills map to one of the four OSLO domains — Offers, Sales, Leads, Operations — so when the framework tells you what to work on, there's a skill ready for an agent to execute it.