
What’s a skill in AI?
A reusable set of instructions that guides an AI through a particular task : summarizing a document, extracting dates, or writing in a specific style. It can include examples, reference material, and scripts.
Think of it as a playbook: it captures how you want the work done.
What’s an agent?
An AI system that works toward a goal by deciding what to do next. It can interpret a request, choose tools, take actions, check results, and adjust its approach.
Think of it as the coordinator: it determines which steps are needed and how to carry them out
So what’s the actual difference?
A skill defines how to perform a task. An agent decides which tasks to perform, when, and what to do with the results.
A skill can guide reasoning as well as execution. The distinction is about responsibility—not “thinking versus doing.”
Give me a real example.
A customer sends a complaint about a delayed order.
A skill might define how to summarize the complaint or draft a response in the company’s tone.
An agent can investigate the complaint, retrieve the order history, apply the escalation policy, and draft a response. It can then request approval or send the response, depending on its permissions.
The tools provide access to systems. The skills guide how the work gets done. The agent coordinates the process.
What if I don’t define skills at all?
The agent can still work using its instructions, tools, and the model’s existing capabilities.
Skills become useful when you want it to follow a repeatable approach. They reduce repeated prompting and make your expectations easier to maintain, review, and test.
They improve consistency; they don’t guarantee it.
So skills are basically prompt engineering?
Prompt engineering is part of it.
A simple skill may be a saved set of instructions. A more complete skill can also include examples, templates, reference documents, scripts, and checks.
The value is in turning a useful approach into something reusable.
When should I just use a skill?
When the task is the same every time. If you can write it down as a clear instruction and it’ll work repeatedly that’s a skill, not an agent.
When do I actually need an agent?
When the path depends on what the system discovers.
Investigating a support incident, for example, may require different logs, queries, and follow-up actions depending on the evidence. An agent can choose the next step as the situation unfolds.
Multiple steps alone don’t require an agent. The need to adapt is what matters.
How does this connect to spec-driven development?
In spec-driven development you write a clear specification before building — what the system should do, how it should behave, what the rules are. Skills are essentially that, but for AI. You’re specifying behavior upfront rather than hoping the model figures it out. The agent is your runtime, the skills are your spec.
So if I have good specs, I have good skills?
Exactly. If you can clearly specify what “done” looks like for a task the inputs, the logic, the expected output that becomes a skill. Vague specs make vague skills make unreliable agents. The discipline is the same.
Do they work together?
Always. Agents are pretty useless without skills they’d have nothing to execute with. And skills alone can’t handle anything that requires judgment. The agent thinks, the skill acts.
Thanks


