AI agents don’t need better prompts. They need better goals
AI chatbots? So 2025. Right now, it’s all about AI agents, and we just got another one, this time from Facebook parent Meta.
Meta Muse is the name of Meta’s new, cloud-based AI agent, and like Gemini Spark, ChatGPT Work, and Claude Cowork, Meta Muse doesn’t just sit in a chatbox. Instead, it’s designed to go out in the world and do things on your behalf.
So … do what, exactly? Check out the “Ideas” section in Muse (you can sign in using your Meta account, and you’ll get free access within certain usage limits) and you’ll find dozens of possibilities, ranging from the usual (make a restaurant reservation, book airline tickets) to the esoteric (contact the proper authorities about a pothole, handle the details of any document you snap a picture of).
These types of goal-seeking abilities aren’t exclusive to Meta Muse; Gemini Spark can do them as well, along with ChatGPT Work, Claude Cowork, and similar agentic AI tools.
Another thing this new class of AI agents share is what they don’t need: prompts, or at least not the kinds of meticulously crafted, behavior-shaping prompts we regularly feed to garden-variety AI chatbots.
Defining outcomes, not behavior
Rather than working in a context-free environment with precious little to go on beside what you tell them, AI agents can browse web sites, absorb documents, and spawn sub-agents to gather data, giving them hard information that “grounds” them in reality.
So what AI agents need from you isn’t so much the “how,” but the “what.” What exactly do you want them to do, and/or not do? Or more specifically, what’s the outcome you’re looking for?
Defining outcomes is new prompt engineering when it comes to AI agents, and you’ll also need to give your agent guidelines to keep it from doing more than you asked for (which, as we know, is a bit of an issue these days).
One of the most important guidelines is defining what “done” means — that is, giving your agent a finish line they can verify, so it doesn’t go overboard and burn tokens pursuing goals you didn’t ask for. If “defining done” sounds familiar, perhaps you’re remembering a prompt I’ve covered that does that very thing.
AI agents also need boundaries, particularly when it comes to buying stuff or accessing local files and directories on your system. Which files can the agent open and/or edit, and which ones are off limits? How much time (and how many tokens) should it spend on a particular task?
Finally, at what point (or points) should the agent stop and check in with you? This element is crucial for tasks where (again) the agent does have purchasing power, or to keep it from making decisions on its own when it runs into an unexpected roadblock.
If that sounds like a lot of elements to juggle within an agentic AI prompt, it is. That’s why I asked Astra, the new ChatGPT supermodel, for a prompt that lets a regular AI chatbot create an “outcome”-oriented prompt for an AI agent.
This prompt (which I’ve pasted below) will kick off a wizard that guides you through the process, asking you what you want the agent to do, what an ideal outcome would be, what constraints the agent should have, and when it should stop to ask you questions.
I tried the prompt on ChatGPT Instant, with a simple, low-stakes goal: Find affordable round-trip airline tickets for the holidays. When asked, I gave a specific boundary: Don’t do anything beyond research.
I then plugged the resulting “outcome” prompt into both Meta Muse and Gemini Spark, and the agents dutifully plowed through reams of airline rates and pinpointed some surprisingly cheap options. Most importantly, neither of the agents went rogue with my credit card.
Here’s the prompt, and remember: Plug it into a standard AI chatbot (like ChatGPT Instant or a “flash” Gemini model) and then hand the result to Meta Muse, Gemini Spark, or another personal AI agent:
Help me turn a rough task idea into an outcome prompt I can give to an AI agent. Guide me through a short wizard; don’t execute the task.
Start by asking: “What would you like an AI agent to accomplish for you? A rough idea is enough.”
Cover four steps, one at a time:
1. OUTCOME
What should be different when the agent finishes, and why does it matter? Turn activities like “research flights” into outcomes like “find three flights that fit my schedule and budget, so I can choose.”
2. DONE
What should the agent deliver, what requirements must it meet, and how should it verify success? Establish a practical stopping point. Separate what it can verify from what it cannot guarantee.
3. BOUNDARIES
What information and tools may it use? What may it change or commit to? Establish time, spending, and privacy limits. Distinguish research from purchasing, booking, messaging, or changing files; don’t assume permission for those actions.
4. WHEN TO ASK
Which decisions can it make independently? When should it ask—for example, missing essential information, conflicting requirements, no suitable results, or actions requiring approval? Specify what it should do while awaiting my reply.
HOW TO GUIDE ME
– Ask one focused question per message, with a few relevant example answers and room for my own response.
– Aim for 4–6 questions total. Reuse information I’ve already provided.
– If I’m unsure, suggest a practical default for me to confirm.
– Distinguish requirements from preferences and suggestions.
– Request only necessary information. Use placeholders for sensitive details; never ask for passwords, payment details, or identification documents.
– Resolve gaps that materially affect success or permissions; label other assumptions.
FINAL OUTPUT
Write one self-contained, ready-to-copy prompt addressed to the agent, including:
– The outcome, context, deliverables, completion checks, and stopping point.
– Hard requirements and softer preferences.
– Access, action, resource, and privacy boundaries.
– Decisions it may make independently, triggers for asking me, and behavior while waiting.
– Instructions to check its capabilities, disclose blocking limitations, and report partial results if completion is impossible within the boundaries.
Keep the prompt proportional to the task and grounded in my answers. Briefly list any assumptions or placeholders I should review afterward. Don’t execute the task.
