Skip to content

← All articles

AI Agents

How to Write Your First Prompt for an AI Agent (Without Coding)

Based on official prompting guidance from Anthropic and Google, consulted on 24 October 2026.

A handwritten brief on a notepad with five labelled parts, goal, context, steps, format and limits, next to a laptop running an AI agent

With a chatbot, a vague question costs you a vague answer. With an agent that opens files, browses sites and builds documents, a vague brief costs you time, and sometimes a mess to undo.

The good news: you don't need code. You need a clear brief.

Anthropic's prompting guide puts it this way: think of the AI "as a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result."

It also suggests a simple test. Show your prompt to a colleague with little context. "If they'd be confused, Claude will be too."

Google's Gemini help pages say something similar. Good instructions cover four areas: the role to play, the task, the context and the format you want.

The five-part brief

1. Goal: what "done" looks like

Say the result, not the activity. "Summarise my invoices" is an activity. "A table of every unpaid invoice from September, with client, amount and due date" is a result.

2. Context: why, and for whom

Anthropic notes that explaining the reason behind an instruction helps the model "deliver more targeted responses." Say who will read the result and what it's for.

3. Steps: when order matters

The same guide recommends giving "instructions as sequential steps using numbered lists" when order or completeness matters. Agents do multi-step work, so this is where you earn your result.

4. Format: exactly what to hand back

File type, columns, length, tone, where to save it. Be specific. Anthropic calls examples "one of the most reliable ways to steer" format and tone, so paste one if you have it.

5. Limits: what not to touch

This part matters most for agents. Say which folders it may use, what it must not delete or send, and when it should stop and ask you.

A ready-to-copy first prompt

Goal: Create a spreadsheet of unpaid invoices from September 2026.

Context: I'm a freelancer. I'll use this to send payment reminders on Friday.

Steps: 1) Open the PDFs in the "Invoices/2026-09" folder. 2) For each one, find client, invoice number, amount and due date. 3) Check the "Payments" sheet and leave out any invoice marked paid.

Format: One spreadsheet, one row per invoice, sorted by due date. Save it as "unpaid-september.xlsx" in the same folder.

Limits: Only use these two locations. Don't email anyone. Don't delete or rename files. If a PDF is unreadable, list it at the end instead of guessing.

Set the guardrails in the tool, too

Your prompt isn't the only control. In Claude Cowork, for example, Anthropic's help centre describes:

If you repeat the same context every time, move it into those standing instructions and keep your prompt short.

Your first week

  1. Pick one boring, low-risk task.
  2. Write the five-part brief.
  3. Run it, then review every line of the result.
  4. Note what went wrong and add one sentence to the brief to prevent it.
  5. Run it again next week.

Go deeper

Our Portuguese-language book Agentes de IA para quem não programa (AI Agents for People Who Don't Program) turns this method into a full delegation plan, with everyday examples and honest limits.

👉 See our books at bookpilot.org/books

Sources

All consulted on 24 October 2026. Product details may have changed since.

  1. Anthropic: Prompting best practices (Claude Platform Docs)
  2. Google: Tips for creating custom Gems (Gemini Apps Help)
  3. Anthropic Help Center: Get started with Cowork