Skills
When instructions get too long, skills let you break them into reusable, shareable pieces that reduce token usage and improve results.
You’ve written great instructions. Your agent knows its role, follows a clear workflow, and handles edge cases. But then you keep adding more. The instructions get longer. You copy sections between agents. You start noticing the agent occasionally ignores rules buried deep in the prompt.
This is the signal that you need skills.
Think of It Like Onboarding a New Hire
Imagine you’re onboarding a new employee. On their first day, you sit them down and explain the essentials: their role, who they report to, the team’s communication style, what matters most. That conversation is short, high-level, and sets the foundation for everything they do.
But you don’t recite the entire employee handbook in that conversation. You don’t walk them through the company style guide, the expense policy, and the incident response playbook all at once. Instead, those documents live on a shared drive. When the new hire needs to write a blog post, you say “check the style guide.” When they need to file an expense report, you point them to the expense policy.
That’s exactly how skills work in Gumloop.
- Instructions are the onboarding conversation. They tell your agent who it is, how to communicate, and what matters most.
- Skills are the reference manuals. Focused, reusable guides for specific tasks that the agent pulls off the shelf only when it needs them.
What Are Skills?
A skill is a focused, reusable set of instructions for a specific task. Instead of cramming everything into one giant system prompt, you break your agent’s capabilities into discrete skills that get loaded only when needed.
| Instructions | Skills | |
|---|---|---|
| Analogy | The onboarding conversation | Reference manuals on the shelf |
| Scope | Identity and general behavior | Specific task or workflow |
| Always loaded | Yes, every conversation | No, only when relevant |
| Shared across agents | No, per agent | Yes, across your workspace |
| Token cost | Paid on every message | Only when activated |
Why Skills Work Better
There are three practical reasons to break long instructions into skills.
Lower token usage. Everything in your instructions gets sent to the model on every single message. It’s like making your new hire re-read the entire employee handbook before answering each question, even if the question has nothing to do with the handbook. Skills only load when they’re relevant, so you’re not paying for context the model isn’t using.
Better results. Long instructions dilute focus. The more rules you pack into a single prompt, the more likely the model is to overlook something. A 200-word skill about email formatting will get better adherence than the same 200 words buried in a 3,000-word system prompt. Shorter, focused guidance means the model pays closer attention to what matters for the task at hand.
Shareable across your workspace. If three agents all need to know how to write reports in your company’s format, you don’t copy-paste those instructions three times. You create one skill and attach it to all three. Update the skill once and every agent gets the improvement. Same idea as having one company style guide that everyone references, rather than pasting formatting rules into every job description.
Instruction or Skill?
Not sure where a piece of guidance belongs? Ask yourself these questions:
- Is it about identity, role, or tone? That’s an instruction. It applies to every conversation.
- Is it a specific process for a specific task? That’s a skill.
- Do multiple agents need the same guidance? Skill. Write it once, share it everywhere.
When in doubt, start with instructions. You can always extract a skill later when things get complex.
How to Create a Skill
The easiest way to create a skill is to let the agent do it for you.
- Have a good conversation. Work with your agent until you get the outcome you want. Maybe it wrote a perfect report, nailed a customer response, or handled a tricky enrichment flow exactly right.
- Ask it to save the skill. Once you’re happy with the result, tell it: “Save this as a skill.” The agent captures the approach it used and packages it as a reusable skill.
- Iterate. Run the skill a few times. When something’s off, refine it. Skills are designed to be edited and improved over time, just like the employee manual gets updated when processes change.
You don’t need to write the skill from scratch or understand prompt engineering. You just need one good result, and the agent does the rest.
When you create a skill, it’s available to every agent in your workspace. Anyone on your team can attach it to their agents. It’s like adding a new manual to the company library rather than just one person’s desk.
Skills in Practice
Say you have a sales prep agent. Its instructions start simple:
- You are a sales preparation assistant.
- When given a company name, enrich the contact, pull CRM history, and draft a briefing email.
That works fine at first. But over time, you add rules about email formatting, report structure, how to handle missing CRM data, which fields to prioritize, what tone to use for different deal stages. The instructions grow to 3,000 words. Your agent starts missing rules. You’re paying for email formatting guidance on every message, even when someone just asks “what’s the status of the Acme deal?”
With skills, you’d break it apart:
- Company Research skill: how to enrich a contact and pull relevant background
- CRM Summary skill: how to read deal history and surface what matters
- Briefing Email skill: your company’s format, tone, and structure for prep emails
The agent’s instructions stay short and focused on identity and role. Each skill loads only when the agent decides it’s relevant. And the Briefing Email skill can be shared with your marketing agent, your support agent, or any other agent that sends formatted emails.