Cheatsheet

Prompt, Skill, Agentor Flow?

Six yes-or-no questions. You never answer more than four.

Prompt, skill, Cowork task, agent and flow aren't five options on one shelf. They answer different questions, and the order you ask them in is the method. Start with the first question, and stop at the first result.

The whole tree at a glance

The same decision as text, for anyone who would rather not click.

1. Does the result get to roll dice?

→ No: a flow (Power Automate)

→ Yes: go to question 2

2. Does it need its own name, knowledge or tools?

→ Yes: an agent, go to question 5

→ No: go to question 3

3. Should it run on its own, on a schedule or an event?

→ Yes: a Cowork task

→ No: go to question 4

4. Will you do this again, more than twice?

→ Yes: a skill

→ No: a prompt

Which coach?

5. Is it for you or a small team, and does it only answer?

→ Yes: Agent Builder

→ No: go to question 6

6. Does it need its own model, code or network?

→ Yes: Microsoft Foundry

→ No: Copilot Studio

The seven results

A prompt

Type it, take the result, move on. Nothing to build, nothing to file, nobody to own it.

One-off work that needs judgement is exactly what chat is for. Anything else is overhead for a task that won't exist next month.

How you'll know it's more than that: You type the same brief for the third time. That's when you move up a step and write a skill.
The common mistake here: Building too early. A prompt you're still changing doesn't belong in a file.

A skill

A text file with a name, a trigger line and the method. Copilot applies it on its own whenever a request matches.

Repeated work with a fixed sequence is a method, not a request. Written down it runs the same way every time, and a colleague gets the same result without you.

What goes in: Trigger line, steps, output format, done criteria, stop rules, and one example of the finished result.
The common mistake here: Knowledge instead of method. Write a wiki page and you get a wiki page with a header.
The moment you share it: It stays yours. Give it a name, a date and one place where the current version lives. A skill without an owner drifts.

A Cowork task

Your work, running without you: on a schedule, or whenever an event happens.

It's your work and it needs judgement, but it should happen while you're not in the room. Neither an agent nor a flow. You, automated.

Skill first, schedule second: A task can use your skills. The skill is the method, the task is the trigger.
The common mistake here: A task that runs every morning runs every morning. Check, when you set it up, that it really should.
What it isn't: Not a substitute for a flow when the result must be identical. Not a substitute for an agent when it needs its own knowledge, tools or name.

An Agent Builder agent

Described in natural language, grounded in content that already exists. Answers for a team.

The agent answers rather than acts, it's for you or a small team, and the users already have Copilot. No code, built in an afternoon, maintained by you.

When it gets too small: The moment it should create a record, call a system or reach people outside your tenant. Then it can be copied to Copilot Studio.
The common mistake here: Skipping it. A Copilot Studio project for a question this would have answered in an afternoon.

A Copilot Studio agent

A managed product: acts in systems, multi-step workflows, integrations, broad audience, lifecycle.

The agent has to do something, not just say something, or the audience is broad. That needs permissions, connectors and operations with someone answerable.

Check the cheaper answer first: Would Agent Builder do? Would a shared skill do, if everyone has Cowork? A Studio agent costs operations, every month.
The common mistake here: An agent for a task that runs the same way every time. That's a flow in an agent costume, pricier and harder to audit.
Settle this before you start: Who owns it, where its knowledge comes from, who sees what, and what happens when it answers wrongly.

A Microsoft Foundry agent

Code-first. A developer team with full control over model, orchestration and runtime.

The agent needs control that low-code can't give: own model choice, own code, own networking, engineering-grade evaluation. Or it's part of a product, not a business workflow.

Check the lighter answer first: If the need still lives inside Microsoft 365 and business workflows, Copilot Studio is faster, under governance you already have.
The clean pattern: Copilot Studio as the front door, Foundry behind it for the heavy logic. Both together, not one instead of the other.
The common mistake here: Starting with the technology instead of the workload. That's the fastest way to get the architecture wrong.

A flow

Power Automate. Fixed sequence, same input, same result, with a log.

When an approval, a figure or an audit depends on it, reproducibility isn't a preference. No language model guarantees you the same result twice.

An agent may still sit in front of it: The agent handles the conversation, the flow handles the step with the audit trail. That isn't an exception, it's the clean pattern.
If the agent should call the flow: Agent flows in Copilot Studio are convenient. Check first who else may change it and who sees the bill. The audited step of a business process belongs in a cloud flow the agent calls.
The common mistake here: Skipping the flow because the agent looked more impressive in the demo.

Prompts are reps. Skills are technique.

Everything above that is equipment, and equipment needs an owner.

FAQ

What's the difference between a prompt, a skill and an agent?

A prompt is a one-time instruction to the Copilot you already have. A skill is a written-down method that the same Copilot applies on its own whenever a request matches, with the access you already have, and you can share it with colleagues. An agent is a thing of its own: it has a name, its own knowledge sources, its own tools and a chat people open to talk to it, whether that's you alone or a whole department. The line between skill and agent isn't who uses it. It's whether it needs its own knowledge and tools.

When is a flow the right choice instead of an agent?

Whenever the result has to be identical on every run and auditable afterwards: approvals, financial steps, compliance records, anything where a number or a signature depends on the outcome. A flow is deterministic; the same input produces the same output, with a run history. Language-model-based agents are not. An agent can still handle the conversation in front of the flow, but the audited step belongs in the flow.

What is a Cowork task?

Work that Copilot Cowork runs for you on a schedule or in response to an event, such as a matching email arriving, without you starting it each time. It's the right choice when the work is yours, needs judgement rather than a fixed sequence, and should happen while you're not there. A Cowork task can use your skills: the skill is the method, the task is the trigger.

Agent Builder or Copilot Studio?

Agent Builder when the agent answers questions from content you already have, for yourself or a small team, and the users already have Copilot. Copilot Studio when the agent needs to act in systems, run multi-step workflows, use integrations, reach a broad or external audience, or when the organisation has to operate it. Start with the lighter one; an agent can grow into Copilot Studio later.

When do I need Microsoft Foundry?

When a developer team needs control that low-code can't give: choosing the model, writing custom orchestration and tools in code, private networking, or engineering-grade evaluation. If the need still lives inside Microsoft 365 and business workflows, Copilot Studio is the faster and better-governed choice. Many organisations use both, with Copilot Studio as the front door and Foundry behind it.

Why does the order of the questions matter?

Because the first question rules things out and the others rank things. If the result has to be identical every time, nothing built on a language model qualifies, so that question comes first. After that, each question is cheaper to answer than the next, and the cheapest answer that works wins. Ask them in a different order and you end up comparing tools that were never peers.

Last updated: 17 September 2026

Part of the guideWhen to Use What