aksel

Agents Are Now the Primary Callers - Skills Are Not Just for Developers

Published on

Very interesting by "AI News & Strategy Daily | Nate B Jones"! This is from the Youtube, so it is "FFA", NOT from the the Substack post, EXTREMELY interesting, BUT is “paywall”, and I bill not break Jones ownership.

AI News & Strategy Daily | Nate B Jones

Agent Readable Skills — Nate B. Jones (YouTube)

Source: AI News & Strategy Daily (YouTube)
Topic: How skills have evolved from personal configuration to organizational infrastructure

Four Major Changes Since October 2025

1. From Personal to Organizational

In October, a skill was something you built for yourself. Now, team and enterprise admins are rolling out skills workplace-wide. They're version controlled, available in the sidebar, and callable from Excel, PowerPoint, Claude, and Copilot.

2. Agents Are Now the Primary Callers

In October, humans were the primary callers of skills. Now the majority of skills are called by agents — because agents can make hundreds of skill calls over the course of a single run. We need to start thinking about our skills as "agent first".

3. Skills Are Not Just for Developers

They're not meant to just live in the terminal. They're meant for your entire business life — and personal life. Anthropic and Microsoft have a partnership to bring skills to Copilot. OpenAI includes skills in their releases. Skills are now an open standard.

4. Skills as Open Standard

People are trading skills like baseball cards. We're learning together how to make a simple markdown file function as an agent-callable context layer. Best practices are discovered collectively, not known in advance.

What Is a Skill?

A skill is a folder with a text file in it.

Requirements: - One file: skill.md - Two parts: Metadata at the top + methodology/instructions below

That's it. A series of plain English instructions that give an LLM context to do something useful with a particular set of inputs in a predictable way.

Common Usage Patterns

The Specialist Stack (Developers)

A developer drops a folder of skills into a project: - One skill turns vague instructions into a PRD - Another decomposes the PRD into GitHub issues - Another helps write tests

The agent in Cursor can invoke skills and just work — without specialist direction from the developer. The methodology is in the files.

Operations Skills (Example: Real Estate)

Texas Paintbrush on X has over 50,000 lines of skills across 50 repositories: - Rent roll standardization - Comps analysis - Cash flow handling - Handoff protocols between team members

Bonus: Writing all this down also helps humans during onboarding. The methodology no longer lives in someone's head — it lives in a repository.

Orchestrator Skills

An "orchestrator" skill analyzes incoming requests and spawns sub-agents based on skills it learns to call. A single high-level request can reliably be distributed to research, coding, UI, or documentation.

Skills Compound, Prompts Evaporate

"Skills compound. Prompts don't."

  • People who have been building skills for 6 months have improved and refined them
  • People who just prompt are copying and pasting the same stuff
  • Skills compound through industry investment in the ecosystem AND your own effort
  • Prompts are Lego 4x4 blocks — you still need specialized blocks to build the castle

How to Build a Skill That Works

The Description Is Critical

Most skills die in the description.

Bad description:

"Helps with competitive analysis"

Too vague. Triggers on anything tangential.

Good description: - Names the document types or artifact types it produces - Includes actual trigger phrases ("analyze our competitors", "who are the players in this market") - States what the output should look like

Anthropic's own guidance: "Skills tend to under-trigger more than over-trigger. Make the description pushy."

Technical constraint: The description MUST stay on a single line. If a code formatter breaks it across multiple lines, Claude will not read it correctly.

The Methodology Body Needs Five Things

  1. Reasoning, not just steps
  2. Give Claude your frameworks, quality criteria, the principles behind your decisions
  3. A skill with only linear procedures is brittle

  4. Reasoning helps Claude generalize

  5. Specified output format

  6. Not "produce a summary" — that's too vague

  7. Is it markdown? Excel? PDF? Which fields or sections?

  8. Explicit edge cases

  9. Everything a human handles with common sense needs to be written down

  10. Claude will NOT just "know" edge cases like an experienced person would

  11. Examples to pattern match against

  12. Put examples in other files in the skills folder

  13. Claude needs to see what "good" looks like

  14. Keep it lean

  15. A short skill that fires reliably beats a long skill with competing instructions
  16. No more than 100-150 lines in the main skill file
  17. 80% of attention on the description
  18. 20% on clear, general-purpose reasoning

Agent-First Skill Design

Failures Are Different Now

When a human saw something drift, they could correct it right there. Now the agent tries to use the skill without a recovery loop — that can get expensive.

You need to start quantitatively testing skill performance to ensure they're ready for agents: - Have a test suite - Run a basket of tests - Change the skill (version number) - See if it does a better job

Four Principles for Agent Callers

  1. The description is a routing signal, not a label
  2. You're telling the agent through the description where to go in the workflow

  3. The description must contain wording that matches the outcome the agent is looking for

  4. Agents need contracts

  5. Think like an API contract
  6. "This is what this skill gives you. This is what it does NOT give you."

  7. A declarative agreement the agent can easily discover

  8. Composability is essential

  9. Don't think of the skill as solving a problem
  10. Think of it as producing output that will be handed off down the chain

  11. Think through the end-to-end experience of agents and skills

  12. Hardwiring = use scripts

  13. If you want to hardwire behavior deterministically, use scripts, not skills
  14. Skills are plain English — agents respect them, but they're not guaranteed
  15. This doesn't mean you're bad at AI — it means you know what AI can do

Team Structure for Skills (Three Tiers)

Tier 1: Standard Skills

Consistent across the organization: - Brand voice - Formatting rules - Approved templates

Can be provisioned widely by enterprise admin.

Tier 2: Methodology Skills

How the organization performs high-value work: - How you structure client deliverables - How senior practitioners get their work done - What would take a new hire months to learn

These often live in senior people's heads and need to get out. This is where the alpha is — if the most skilled product person on the team has a skill, the rest of the PM team benefits.

Tier 3: Personal Workflow Skills

Things you do under the desk for your day-to-day.

Warning: Don't just keep them on your laptop. If you're sick or on vacation, someone will wish they could use your tool.

The Missing Skills Library

There are plenty of skills for developers — TypeScript, Angular, Git automation.

What's missing: Domain-specific skills for knowledge work: - Competitive analysis - Financial model review - Deal memo drafting - Research synthesis - Meeting synthesis

Simon Willison wrote in October that skills were "maybe a bigger deal than MCP." Because they're just text files that travel everywhere.

Practical Tips

Beginner:

Look at something you do repetitively (once a week, twice a week) and ask: "Can I turn this into a skill?" Ask your AI to help you create a skill.md from your conversations.

Intermediate:

  • Are you thinking through the handoffs?
  • Are you structuring skills as contracts?
  • Are you agent-first?

Team/Enterprise:

Use the tier model: - Is this an individual skill? - Does this represent the expertise of the best person on the team? - Is it a brand standard that needs to be everywhere?

Key Insight

"Skills compound. Prompts evaporate."

Skills represent a stored record of successful workflow execution that an agent or human can follow. If you continually evolve it as you get better, you'll have a remembered way for future smarter agents and your team to execute that skill.

You free yourself from copy-paste hell. Prompts evaporate when the conversation ends. Skills are what persists.