The Problem With Prompts That Live Nowhere
Most teams using AI today are operating at what we call Level 1: someone figures out a prompt that works, uses it for a week, loses it, and starts over. The knowledge is real, the effort was genuine, but nothing was captured. You end up rebuilding the same capabilities over and over, and every new team member starts from scratch.
That is not a technology problem. It is an organisational design problem. And the solution — a structured skills library — is more achievable than most people realise.
What a "Skill" Actually Is
In the agent context, a skill is a modular, reusable capability. Think of it like a well-written procedure — except instead of instructing a human, it instructs an AI. A skill has a clear purpose, defined inputs, predictable outputs, and enough context that someone unfamiliar with the original use case can pick it up and apply it.
A prompt like "write me a follow-up email" is not a skill. A skill says: given a contact name, a meeting summary, and an intended next step, produce a 150-word follow-up email in this voice, referencing these specific elements. One is a starting point. The other is a reusable asset.
The open-source community has started formalising this. The AGENTS.md standard — adopted across more than 60,000 GitHub repositories — gives agents a structured way to understand purpose, tools, and constraints. The Model Context Protocol (MCP) is doing the same for tool connections. These are infrastructure decisions validating what we've told clients for two years: the organisations that build structured, documented skills now will have a compounding advantage over those that keep winging it.
The Five-Level Maturity Model
Level 1 — Ad-hoc prompting. You type something into ChatGPT or Claude, get a useful result, and move on. No capture, no repeatability, no institutional memory.
Level 2 — Saved prompts and templates. You start saving prompts that work — in a Notion page, a shared doc, a folder. Meaningful progress, but saved prompts without context go stale. Someone opens the doc six months later and has no idea what it was designed to do.
Level 3 — Documented skills with inputs and outputs. This is where real leverage begins. Each capability is documented with its purpose, required inputs, expected outputs, the platform it was tested on, and known limitations. Skills are transferable — a new team member can pick one up and run it on day one.
Level 4 — Version-controlled, portable skills. Skills are stored in version control with change history and the ability to roll back. Critically, they're written to work across platforms. ServiceNow recently opened its skills platform to third-party integrations, and similar moves are happening across enterprise software.
Level 5 — Orchestrated skill stacks. Skills chain together. An intake skill feeds a research skill, which feeds a drafting skill, which feeds a quality-review skill. You're building pipelines, not just prompts.
Getting From Level 1 to Level 3 in 30 Days
Week One: The Audit
Ask every person on your team to list tasks they do more than twice a month involving writing, summarising, researching, formatting, or communicating. From that list, identify the five to eight with the highest frequency or the most inconsistency in quality. A team of ten people typically surfaces 20 to 40 candidates in under two hours.
Week Two: Document Before You Automate
Before writing a single prompt, document what each skill does in plain language. We use a simple template: Purpose (one sentence), Inputs (the data needed), Output (format and length), Quality criteria (what good looks like), and Known limitations. This takes about 20 minutes per skill. It feels slow. It is the difference between a library that lasts and a folder of forgotten prompts.
Weeks Three and Four: Build and Test
Write the prompts for your three best-documented skills. Use structured output formats (JSON, markdown tables) because they transfer more reliably than prose instructions. Test each on at least two platforms. A skill that works on one platform only is a dependency, not an asset.
By the end of week four, your goal is five documented, tested, working skills in a shared location. That is Level 3. It's not massive. But it's real — and it compounds from there.
What Makes Skills Last
The organisations getting sustained value treat skills as living documents — reviewed quarterly, with assigned ownership. They resist the temptation to build everything at once. Starting with five excellent skills beats fifty mediocre ones every time.
Your skills library is not just a folder of prompts. It is institutional knowledge, made reusable.