AI prompts vs skills vs workflows: which one to use

This is taking longer than expected.

Try againBrowse the library
AI Blocks

Blog Get started with AI

AI prompts vs skills vs workflows: which one to use

Comparing AI prompts vs skills comes down to how much of a job you want to save for next time. A prompt is a single message, while a skill stores detailed instructions for a repeated job. A workflow connects several skills in sequence, and an agent works through the steps independently, using tools.

AI prompts vs skills: the short answer

When people compare AI prompts vs skills, they are really asking how much of a job they should save for next time. A prompt is one message you paste into a chat to get one result. A skill is a saved set of instructions, with steps and checks, for a job you do repeatedly. A workflow joins several skills in a fixed order, so that each step hands its work to the next. An agent is an AI that works through the steps independently, using tools such as files or web search.

The practical rule is to start with the smallest option that does the job, and to move up a level only when the smaller one stops being enough.

KindWhat it isBest for
PromptOne message you paste into your AI.A one-off job that fits in one message.
SkillA saved file of steps, rules and checks for one job.A job you do every week, the same way.
WorkflowSeveral skills in a set order.A bigger job with several parts.
AgentAn AI that picks its own next step and uses tools.A job you want done while you do something else.
SpecialistA ready-made AI role for a whole area of work.Handing over a role, like bookkeeping help, with its methods already written.

The rest of this guide explains each option in more detail, shows the same job done four different ways, and gives you a decision table for your own work.

What is a prompt?

A prompt is the message you type or paste into ChatGPT, Claude or Gemini to request one specific thing, such as ten subject lines or a document summary.

Prompts are quick and need no preparation, but their weakness is memory, because each time you paste the whole thing again and correct the same problems again. A good prompt has six parts: role, context, task, inputs, format and checks. Our guide on how to write ChatGPT prompts for work walks through each part.

A prompt is the right choice when the job is unfamiliar, when you need it only once, or when you are still deciding what a good result looks like.

What is an AI skill?

An AI skill is a set of instructions saved in a file, usually called SKILL.md, that tells your AI how to do one particular job. It covers what to ask you for, the steps to follow, the rules to keep and the checks to run before it hands work back. A skill can also hold supporting files, such as a template or a checklist, in the same folder.

The important difference from a prompt is that a skill is permanent, so you write it once, improve it after each run and reuse it indefinitely. In apps that support skills, your AI reads each skill's short description and loads the full instructions only when a request needs them.

Where skills work

  • Claude: the Claude apps and Claude Code both support skills, and our guide to using Claude skills explains the installation steps.
  • An open format: the skill folder format is published as an open standard called Agent Skills, so other compatible tools can read the same files.
  • Gemini: Google now offers skills in the Gemini app, and its help centre says Gems on personal Google accounts start moving to skills in November 2026 (see Google's Gems help page).
  • Any chat AI: even where skills aren't built in, you can attach the SKILL.md file to a conversation and ask your AI to follow it.

A skill is the right choice when you do the same job at least every week or two and find yourself giving the same corrections each time.

AI workflow vs agent: what is the difference?

A workflow is a set of steps in a fixed order: perhaps one step writes a brief, the next drafts the copy and a third checks it. Each step receives what the previous step produced, so you know in advance exactly what will happen.

An agent makes more decisions independently. You give it a goal and access to tools, and it chooses what to do next, such as reading a file, searching, drafting or checking. A chat AI only writes, while an agent can also act, for example by opening files or using other applications you connect.

Anthropic, the maker of Claude, draws the same line in its guide to building effective agents. Workflows follow paths set in advance, whereas agents direct their own steps and choose their own tools. Its advice is to find the simplest solution that works and to add complexity only when it is needed. Workflows give you predictable results on well-defined jobs, while agents suit jobs where the next step depends on what they discover along the way.

What are specialists and minds?

These are the two other terms you'll see on AI Blocks, and both describe ready-made setups you can download.

  • A specialist is a ready-made AI role, such as a bookkeeper or a designer. It comes with a job description, quality checks, permissions (what it may do alone and what needs your approval) and training notes. You download it and give it to your AI, and you can browse them under hire an AI specialist.
  • A mind is the thinking style of a well-known expert, drawn from their public work, so that your AI can examine a problem the way they might. It is an interpretation of their published ideas rather than the person, and you can explore them under minds to think with.

Think of a specialist as hiring someone for a role, and think of a mind as asking how a particular expert would approach your problem.

Which should you use? A decision table

Find the line in this table that best describes your situation, then read across to see which option fits and why.

If you...UseWhy
need something once, right nowA promptNo setup. Paste, fill in the brackets, send.
do the same job every weekA skillWrite the steps and checks once, then reuse them.
have a job with several parts that feed each otherA workflowEach step gets exactly what it needs from the one before.
want it done while you do something elseAn agentIt works through the steps with tools, and stops to ask when it should.
want a whole role covered, not one taskA specialistThe role's methods, checks and limits come already written.
want a fresh way to think about a hard choiceA mindYour AI reviews the problem the way that expert might.

Three questions to ask

  1. How often do I do this job? A one-off job suits a prompt, while a weekly job usually justifies a skill.
  2. How many separate parts does it have? A single-part job suits a prompt or skill, while several parts that pass work along suggest a workflow.
  3. Do I need to supervise each individual step? If you do, run it yourself in a chat, but if you don't and the job is low risk, consider an agent.

A worked example: Basecamp's Shape Up pitch, four ways

Basecamp is project management software made by the company 37signals, whose team published a free book by Ryan Singer, Shape Up, about how they plan product work. In Shape Up, a project starts as a written "pitch" with five parts. They are the problem, the appetite (how much time it is worth), the solution, the rabbit holes (risks to call out) and the no-gos (exclusions).

Imagine your own team wants to write pitches this way, and here is how the same job looks at each level. It is an illustration written for this article, and we are not suggesting that 37signals uses AI to write pitches.

As a prompt

For your first pitch, you paste your notes and ask your AI to write a pitch with the five Shape Up parts, using only the information in your notes.

As a skill

After a few pitches, you notice that you keep making the same corrections, because the appetite is often missing and the solution is too detailed. So you save a reusable skill, and here is a shortened example of the file:

---
name: shape-up-pitch-writer
description: Writes a project pitch with the five Shape Up parts from rough notes. Use when someone asks for a pitch or a project proposal.
---
# Shape Up pitch writer

Ask for: the notes, the appetite (small batch or big batch) and who reads the pitch.

Steps:
1. Write the problem in the reader's words, with one real example.
2. State the appetite. If none is given, ask. Never guess.
3. Sketch the solution at a high level. No pixel-level detail.
4. List the rabbit holes and how to avoid each one.
5. List the no-gos.

Checks before you answer:
- All five parts are present.
- Nothing is added that is not in the notes.
- The solution fits the appetite.

As a workflow

Later, the pitch becomes one step in a bigger, multi-part job. The first step turns customer requests into a problem statement, and the second writes the pitch with the skill above. The third step reviews the pitch for missing rabbit holes, and the fourth drafts a short note for the team that will build it. Each step automatically passes its result to the next.

As an agent

Finally, you give an agent a folder of customer requests and the goal "draft pitches for the three most requested problems". It reads the folder, groups similar requests, runs the workflow for each group and stops before sharing anything, so that you can review the drafts yourself.

Notice how each level preserves what the level below it learned. The skill holds the corrections from your early prompts, the workflow uses the skill, and the agent runs the workflow.

Two prompts to try this on your own work

Use the first prompt below to decide which kind of setup a job needs, and the second to turn a prompt you keep reusing into a skill.

Prompt: pick a prompt, skill, workflow or agent

You are an adviser who helps small business owners set up AI for their work.

Context: I want help with this job: [describe the job]. I do it [how often]. It takes me about [time]. It needs these things from me: [the material, for example notes, a price list, past emails]. A mistake would cost [low, medium or high].

Task: Tell me whether this job needs a prompt, a skill, a workflow or an agent.

Use these definitions:
- Prompt: one message for one result.
- Skill: saved steps, rules and checks for a job done often.
- Workflow: several steps in order, each passing work to the next.
- Agent: an AI that chooses its next step and uses tools.

Format:
1. Your pick, in one line.
2. Three reasons, tied to what I told you.
3. What I would need to set it up.
4. One sign that I should move up a level later.

Checks: Pick the simplest option that does the job. If a mistake would be costly, say where a person should approve the work. If you need more facts, ask me first.

Prompt: turn a prompt you reuse into a skill

You are an editor who turns prompts people reuse into skill files.

Context: I use the prompt below often. I keep making the same corrections to the answers. The skill will be saved as SKILL.md and used in [Claude, ChatGPT, Gemini or another AI].

Inputs:
PROMPT START
[paste the prompt you reuse]
PROMPT END
CORRECTIONS START
[list the fixes you keep making, one per line]
CORRECTIONS END

Task: Write a SKILL.md for this job.

Format:
- Start with a front matter block between two lines of three dashes, with "name" (lowercase words joined by hyphens) and "description" (what it does and when to use it, under 200 characters).
- Then sections: what to ask me for, numbered steps, rules, and checks to run before answering.
- Turn every correction into a rule or a check.

Checks: Keep it under 60 lines. Do not add steps I did not ask for. List anything you were unsure about at the end.

Common mistakes, and what to do instead

MistakeDo this instead
Building an agent for a job you have never done with a promptDo it by hand with a prompt first. You can't automate a job you can't describe.
Writing one giant skill for "all my marketing"Make one skill per job. A skill for ad headlines and another for product pages.
Joining steps into a workflow before each step works aloneTest each skill on its own, then join them.
Letting an agent send, post or spend without askingStart with read-and-report jobs. Add an approval step before anything goes out.
Never updating a skillWhen you correct the result, add the fix to the skill as a rule.

How AI Blocks uses these words, and where it helps

AI Blocks is a free library that uses these same terms for six kinds of item: prompts, skills, workflows, agents, minds and specialists. You copy a prompt, or download a skill, workflow, agent or specialist, and then use it in your own AI, because AI Blocks doesn't run the AI itself. The how it works page has the complete guide, with examples of every kind of item.

  • The library: search by job, or browse one kind of item, such as prompts or skills.
  • Plan a job: describe a goal and receive a step-by-step workflow you can download.
  • Reusable AI skill writer: paste a job you repeat and get a skill draft with a trigger line, steps and three test prompts.
  • Skill writer and reviewer: write a skill from a job description, or review one you have, with a clearer trigger line and safety notes.
  • Support reply and help-page writer: an example workflow that sorts support tickets, writes and checks replies, and then turns confirmed fixes into help pages.
  • Hire an AI specialist: choose from over 1,000 ready-made role setups that you can download.
  • Train your own AI agent: a step-by-step guide and free starter pack for teaching an agent one job your way.

For more detail on agents, read AI agents for small business, and if you want to build a workflow without code, see how to build an AI workflow without code.

Questions people ask

What is the difference between a prompt and a skill?

A prompt is one message you paste each time, whereas a skill is a saved set of instructions that your AI can reuse for the same job. Prompts suit occasional one-off jobs, and skills suit jobs you repeat regularly.

Are Claude skills the same as custom GPTs?

They overlap considerably, but they aren't identical. A custom GPT is a version of ChatGPT set up with its own instructions and files. A skill is a folder with a SKILL.md file that your AI loads, alongside your other skills, when a job requires it. Both approaches save you from repeating the same instructions.

Is an AI workflow the same as an automation?

Not quite. An automation tool such as Zapier or Make moves data between applications when something happens. An AI workflow is a sequence of AI steps, such as research, drafting and checking. You can combine the two, since an automation can start an AI workflow.

Do I need to code to use skills or workflows?

No, because skills and workflows are plain text files that you can read and edit, and you simply give them to your AI and ask it to follow them. Some jobs do need tools, such as file access, which your AI app may or may not offer.

When should I use an agent?

An agent makes sense when a job has several steps, you don't need to watch each one, and a mistake would be easy to catch. Start with agents that only read and report, and make sure they ask you before they send, post or spend anything.

All articles and guides