|
Almost Timely News: 🗞️ Highly Opinionated Method of Building AI Agent Skills (2026-10-04)
I have strong opinions about Agent Skills
|
| |||||||||||||||
Almost Timely News: 🗞️ Highly Opinionated Method of Building AI Agent Skills (2026-10-04) :: View in Browser
 |
The Big Plug
✍️ Enroll in my newest course, Introduction to Deep Research With AI.
👍 This issue features a new paid download, the The Highly Opinionated Agent Skill Fixer Download!
Content Authenticity Statement
100% of this week’s newsletter was made by me, the human. Learn why this kind of disclosure is a good idea and might be required for anyone doing business in any capacity with the EU in the near future.
Watch This Newsletter On YouTube 📺
Click here for the video 📺 version of this newsletter on YouTube »
Click here for an MP3 audio 🎧 only version »
What’s On My Mind: Highly Opinionated Method of Building AI Agent Skills
This week, let’s talk about building Agent Skills for level 3 and 4 AI systems. If you’re not familiar or you need a refresher, Agent Skills (hereafter Skills with a capital S) are a way to convert a repeatable process into a command you or an AI can use to automate the process.
There are no shortage of guides about how to build Skills, but almost all of them are predicated on the same basic concept: have AI do repeatable tasks. The reason why this issue is titled “Highly Opinionated Skill Building” is because I have strong opinions about the state of most Skills, namely that they’re not great.
So today, we’ll look at my particular point of view. I’m not saying it’s right or wrong beyond a certain point, but it works for me and if you share some of my values, it might work for you too.
Part 1: A Brief Review of AI Levels
I mentioned level 3 and 4 systems above. There are, from my point of view, 5 levels of AI systems, and each level builds on the previous. Here’s what they are, in short:
Level 1: Done by you. Pure chat systems. These are the web-based UIs we all know - ChatGPT, Gemini, Claude, Copilot, etc. In these, you’re doing 100% of the work, prompting, reviewing, copy pasting in and out. This was most people’s AI experience from 2023-2024.
Level 2: Done with you. Systems with standard operating procedures. These are systems that have stored procedures in them, like ChatGPT Custom GPTs, Gemini Gems, Claude Projects, Copilot Agents (which are not really agents) etc. where you store prompts in reusable mini-apps as standard operating procedures. You’re still doing 60% of the work, but you at least don’t have to copy paste prompts much any more, at least not for routine things.
Level 3: Done for you. Systems that take project plans. These are agentic AI systems, systems like Claude Cowork, Claude Code, ChatGPT Work, Gemini Spark, Google Antigravity, Microsoft Copilot Cowork, etc. You give these tools a proper project plan and they go off and do the thing and come back with a result. You’re doing 10-20% of the work, while the system does the heavy lifting and has connectors to other systems. These systems use Skills and Plugins (plugins are bundles of Skills).
Level 4: Done without you. Systems that take job descriptions. These are semi-autonomous agents like Grok Bot, Meta Muse, Hermes Agent, OpenClaw, Claude Managed Agents, OpenAI Dots etc. Give these tools a job description composed of things like project plans and the appropriate tools, and they run almost completely by themselves.
Level 5: Done anticipating you. Systems that have charters. These are fully autonomous systems that have swarms of agents, behaving like an agency with a team of people and a persistent memory. It’s where level 4 systems are evolving to go. these are systems that are truly proactive in that they anticipate, based on past patterns of data, what it is you want them to do now.
In these 5 levels, each level builds on the previous level; you can’t build level 4 systems without having working levels 1-3 infrastructure, in the same way that you can’t bake a pizza if you have never baked anything before.
Part 2: Non-Negotiables
The first and foremost thing that any Skill has to do is conform to the Agent Skills specification. Anthropic built this originally, a year ago, and made it an open framework. Pretty much every major AI platform has adopted it since then, so much so that companies are deprecating their Level 2 platforms in favor of skills.
You can read the standard at AgentSkills.io, but the short version is that a Skill has at a minimum a skill definition file called SKILL.md, plus can have things like templates, assets, references, and scripts (code).
You can immediately make any skill better off the bat by telling your favorite AI agent that at a bare minimum, the skill should conform to the spec. That’s as simple as a prompt like this:
Today you’re aligning software to a specification. Take the Agent Skill here [attached or uploaded] and ensure it fully complies with the Agent Skill specification at https://agentskills.io/specification. Where it doesn’t comply, audit what it would need to do to comply, and tell the user the recommended edits to the Skill to reach compliance. Ask me up to 20 clarifying questions until you have enough information to succeed at the task.
That alone will solve a fair number of problems. For example, I’ve seen Skills that have thousands of words in them, entire narratives, when the spec says that you can have a maximum of 1,024 characters.
What’s really important about Skills, though, are those “optional” components:
Templates: a directory of file examples for Skills that have deterministic outputs where it has to have the same format every time
Assets: a directory of components that the Skill should use or re-use, like logos, graphic files, style sheets, etc. where AI shouldn’t be winging it