Ask for the same thing twice? Turn it into an agent.

Your best agents are already hiding in your chat history. They are the requests you type again and again: the monthly report, the post in your tone, the competitor check before a pitch. Every time, someone explains the context from scratch, and every time the quality depends on who asked and how well.

A better prompt does not fix that. Writing the work down once does. Here is how.

“Building a skill for an agent is like putting together an onboarding guide for a new hire.” · Anthropic

Agent or skill?

First decide what you are writing down.

• An agent is a role that owns a kind of work from request to result: a report builder, a content writer.
• A skill is a method that several roles can use: human writing, translating meaning instead of words.

A quick test: if you would hire someone for it, write an agent. If you would hand people a checklist or a style guide, write a skill.

Six signals that work should be written down

• You have typed the same request, or nearly the same, at least twice.
• You paste the same context every time: brand rules, a template, a list of competitors.
• The result depends on who asks. One colleague gets a good answer, another a mediocre one.
• You correct the same mistakes again and again.
• Only one person really knows how it is done.
• You could explain the steps to a new colleague.

One counter-signal. OpenAI’s practical guide to building agents points them at work that needs judgement, rests on rules that are hard to maintain, or relies on unstructured data such as documents and language. Where none of that applies, “a deterministic solution may suffice”. If a spreadsheet formula does the job, keep the formula.

Write down seven things first

Before any instructions, answer these. They become the backbone of the agent or skill.

1. Goal. What should the result achieve, and for whom? One sentence.
2. Inputs. What does the person give when they ask: a period, a client, a channel? And what should it ask for when something is missing?
3. Sources. Where the facts come from: which knowledge, which data, which tools. Also say what does not count as a source.
4. Format. What the result looks like: file type, structure, length, language.
5. Quality bar. How you will know it is good. Anthropic’s guidance on success criteria asks for criteria that are specific, measurable, achievable and relevant. “Every ratio recalculated from the raw numbers” passes that test; “a good report” does not.
6. What it must never do. Invent figures, write to outsiders, publish without approval. Write the limits down as plainly as the steps.
7. Who approves. Name the steps that need a person: sending, publishing, spending money. OpenAI recommends that high-risk actions “trigger human oversight until confidence in the agent’s reliability grows”.

Do not start from a blank page. The same guide advises building instructions from what you already have: “existing operating procedures, support scripts, or policy documents”.

For a complete example of a method written down, look at the Website Builder: seven steps from goal and market to delivery, a quality bar in numbers, a list of what it never does, and an Owner or Admin who approves before anything goes live.

Test it before you share it

1. Collect three to five real requests from the past, including one awkward one: missing input, an unusual period, data that contradicts itself.
2. Run them without the agent first and note what goes wrong. That is your baseline.
3. Run them with the agent and hold each result against your quality bar.
4. Open the actual output, not just the summary in the chat.
5. Let someone other than the author try it.

This follows Anthropic’s advice for skills: create evaluations before writing extensive documentation, at least three of them, and measure performance without the skill first, so you can see what it adds.

Improve it after use

The first version is never the last. A few habits keep it sharp:

• Fix it once, in the instructions. If you correct the same thing twice, it becomes a rule. That is the title of this article, one level down.
• Watch how it is used. Anthropic advises paying attention to how the model actually uses a skill, and iterating “based on observations rather than assumptions”.
• Ask what went wrong. When a run goes off track, Anthropic suggests asking the model to reflect on what went wrong. Write the lesson into the method.
• Keep it short. Only add what the model does not already know, and cut rules nobody needs.
• Re-run your tests after every change, and when you switch models: the effect of a skill depends on the model underneath.
• Name an owner. One person keeps the agent or skill up to date and decides on changes.

How this works in Zaela

In Zaela, this is how the company brain grows: everything your team repeats becomes an agent or a method the whole team has, and every lesson becomes an agent, a method or memory. Everything the AI uses, you can also see, adjust and add to in the web app. See how it works.