Systems•6 min read•beginner•Updated Oct 10, 2026

Prompt, Skill, Workflow, Agent: Telling the Layers Apart

Skill vs prompt vs agent: a prompt is input for one run, a skill is a reusable versioned package, an agent picks its own next step. Where system prompts fit.

SYSTEMS
Five layers. One slice. Stop climbing.
PRUNINGMYPOTHOS.COM
promptsystem promptskillworkflowchoiceAGENTit's a cake.

Someone says the team is building an agent. Someone else says it is really just a workflow. A third person points out it is mostly a long system prompt. All three can be describing the same system correctly, because these words name different parts of it.

The argument is usually not about definitions. It is about which part somebody is looking at.

Skill vs prompt

A prompt is what goes in for one run. A skill is what you write down once so the same kind of run can be repeated: the instructions, plus what it takes, what it returns, what counts as success and what happens when it fails. A skill usually contains a prompt; it is not a longer one. A saved prompt with a name is still a saved prompt until it declares those things.

Prompt vs agent

A prompt is input to a single run. "Agent" describes who decides what happens after that run: when the model chooses the next step, that part of the system is agent-like. An agent-like system still sends the model a prompt on every step, so the two are not alternatives.

The five words

A prompt is the input package supplied to the model for a single run. Not only an instruction: it can carry context, examples, retrieved material and the question itself. Whatever goes in for this run is the prompt. What that does and does not achieve is covered in prompting is not the skill you think it is.

A system prompt is system-level instruction supplied with the model's context to shape behaviour for a task or interaction. Applications and providers differ in how those instructions are represented and carried across calls. Its job is to set conditions that hold regardless of what is asked. It has its own page, because the interesting part is what it costs and what it cannot do.

  1. A prompt

    • The input package for a single run
    • Instruction, context, examples, retrieved material, the question
  2. A system prompt

    • System-level instruction
    • Sets conditions that hold regardless of what is asked
  3. Applications differ in how those instructions are carried across calls.

A prompt is the input package for one run. A system prompt sets conditions that hold regardless of what is asked.Two of the five words, which get used as one.

A skill, in the sense this publication uses the word, is a reusable, versioned package of instructions and declared operating constraints for a repeated task. It may define inputs, outputs, tool scope, success criteria, failure conditions and escalation. Other platforms use "skill" differently, sometimes for a tool, sometimes for a plugin, sometimes for a prompt template with a name. This is working vocabulary for explaining PMP's own architecture, not a claim about the industry.

A workflow has a sequence primarily defined by software or process logic, even when individual steps use models. Agent-like behaviour appears when the model receives bounded discretion over which action or step happens next. Both of those definitions belong to AI agents vs AI workflows, and this page adopts them rather than restating them in different words.

They are not stages

The most common mistake is reading these as a ladder, with prompts for beginners and agents for advanced teams.

These terms name different responsibilities, and one system usually contains several of them at once. A single mature system can have a system prompt setting conditions, three skills packaging its repeated tasks, a workflow sequencing them, and one step where the model picks among a few options. Nothing in that list is more advanced than anything else.

Reading them as stages produces a specific bad decision: teams add agent-like freedom because it sounds like progress, when a fixed sequence would have been more reliable and easier to debug.

  1. Read as a ladder

    • Prompts, for beginners
    • Agents, for advanced teams
  2. One system, several parts at once

    • A system prompt setting conditions
    • Three skills packaging repeated tasks
    • A workflow sequencing them
    • One step where the model picks
  3. Reading them as stages adds agent-like freedom where a fixed sequence would have been more reliable.

These terms name different responsibilities, and one system usually contains several of them at once.Nothing in that list is more advanced than anything else.

Telling them apart

Four questions separate them in practice:

What is supplied for this run? The prompt.

What is present regardless of the run? The system prompt.

What is packaged for reuse across runs? A skill.

Who decides what happens after this step? If the program decides, it is a workflow. If the model decides, that part is agent-like.

  1. What is supplied for this run?

    The prompt

  2. What is present regardless of the run?

    The system prompt

  3. What is packaged for reuse across runs?

    A skill

  4. Who decides what happens after this step?

    If the program decides, it is a workflow. If the model decides, that part is agent-like

Four questions separate the five words in practice.The last question is the one that changes how much can go wrong.

The last question is the one that changes how much can go wrong, which is why it gets a page of its own rather than a row in a table here.

A word about "runtime"

The word gets used loosely, and one usage is worth correcting because this page previously had it backwards.

The runtime is the host program: the thing that holds credentials, calls tools, enforces limits and decides what is permitted. It is not the agent. When a model is given discretion over the next step, the runtime is the layer that constrains that discretion. Calling the agent "the runtime actor" inverts the relationship and makes the controlling layer disappear from the description.

  1. The runtime: the host program

    The model's discretion

    • Which step happens next, among what it is allowed
  2. What the runtime holds

    • The credentials
    • The tool calls
    • The limits, and what is permitted
  3. The runtime: the host program

    When a model is given discretion over the next step, the runtime is the layer that constrains that discretion.

When a model is given discretion over the next step, the runtime is the layer that constrains that discretion.Calling the agent the runtime inverts the relationship.

Where mistakes show up

Three recur:

A saved prompt called a skill. Naming it does not make it reusable. What makes it a skill in this sense is declaring what it takes, what it returns, what counts as success and what happens when it fails.

A workflow called an agent because it makes several tool calls. Tool use is a separate question from step selection. A workflow can call many tools and still have every branch written in advance.

Task logic living in the system prompt. A system prompt accumulating rules about specific tasks is usually a skill that has not been named yet, and the cost is that nobody reviews it.

  1. What it gets called

    • A saved prompt: a skill
    • Several tool calls: an agent
    • Task rules: the system prompt
  2. What it actually is

    • A skill once it declares inputs, outputs, success, failure
    • A workflow, with every branch written in advance
    • An unnamed skill nobody reviews
  3. A system prompt accumulating rules about specific tasks is usually a skill that has not been named yet.

Naming it does not make it what it is called.Three mistakes recur.

What this changes

When someone asks whether to build an agent, the more useful question is which layer the problem actually sits in. Most of the time the answer is a skill that needs defining or a workflow that needs writing down, and the agent question turns out not to be the decision in front of them.

Where this helps, and where it stops

What it is
A map of five terms, naming what each covers and pointing to the page that owns each definition.
What it does not guarantee
A definition of agent, workflow or tool use, which are settled elsewhere, and not a maturity model.
When the distinction matters
A team is arguing about whether the thing they are building is an agent.