Prompts vs Skills vs Workflows vs Agents — when to use which

If you are using AI for real work, you will eventually hit a question that the chat window cannot answer well: "I do this kind of task every week — is there a better way than copy-pasting the same prompt?" The answer is yes, and there are four levels of "better." Each one trades flexibility for repeatability in a different way.

This guide is the decision framework: when a prompt is enough, when you want a skill, when a workflow earns its complexity, and when an agent actually makes sense.

The four formats, in one paragraph each

A prompt is a single text string you send to the AI. One input, one output. You write the prompt, paste your context, and you are done. Prompts are flexible and zero setup. They do not remember anything between conversations.

A skill is an installable instruction file (typically SKILL.md) that the AI loads automatically when the task matches. Once installed, the AI behaves consistently for that task type without you having to remember to paste the prompt. Skills are persistent and consistent, but still single-step.

A workflow is a step-by-step process — typically 4-7 steps — that the AI takes you through with decision points, copy-paste prompts at each step, and warnings. You follow it manually. The AI does not run the steps for you; you do. Workflows make multi-stage work repeatable without giving up control.

An agent is a system where AI runs multiple steps autonomously and decides what to do next at each step. You give it a goal; it plans, acts, observes, and iterates without your input between steps. Agents trade your control for speed, and they fail in less predictable ways than the other three.

That paragraph is the whole framework. The rest of this guide is the "when does each one actually win" question.

When a prompt is enough

A single prompt is the right choice when:

The task is one-off or rare.
The context is short enough to paste each time without it being painful.
You want full control over the output and might tweak the wording every time.
You are still figuring out what "good output" looks like for this task.

Most AI use should stay at the prompt level. People over-engineer. They build a skill or workflow for something they only do twice a year. That is wasted effort.

The signal that you need to graduate from a prompt: you have copy-pasted the same prompt structure more than five times in the last month.

Browse our prompt library →

When a skill earns its complexity

A skill (a SKILL.md file the AI loads automatically) is the right choice when:

You do this task often enough that retyping the prompt is friction.
The output format should be consistent across sessions.
You want the AI to remember the rules without you re-explaining them.
The task is still single-step — one input, one output — but you want consistency.

Example: a sales rep who writes follow-up emails after every call. The skill says "always summarize next steps in 3 bullets, never include meeting times the prospect did not commit to, default to 80 words or fewer." The AI applies this every time, in every conversation, without the rep having to re-paste the rules.

Skills are the best leverage for recurring solo work. They scale well across sessions. They do not coordinate multi-step work.

Browse our Claude Skills library →

When a workflow makes sense

A workflow — a 4-7 step playbook with copy-paste prompts at each step — is the right choice when:

The work is multi-step and the steps depend on each other.
The output of one step needs to be evaluated before the next step runs.
There are decision points where the right answer depends on what came back from the AI.
You want the speed of AI inside a process you stay in control of.

Example: a cold outreach workflow that goes (1) research prospect → (2) draft email → (3) review for tone → (4) personalize opener → (5) send. Each step has its own prompt. Step 2 uses the output of step 1. You check step 3 before continuing to step 4.

Workflows are the right tool when:

A single prompt would be too long or too complex.
You need a human checkpoint between steps.
The "right answer" varies based on intermediate results.

You stay in the loop. The AI helps at each step. Nothing happens without you running it.

Browse our workflow packs →

When (and when not) to use an agent

An agent runs multiple steps without you between them. Goal in, completed work out. The agent decides what to do next at each step based on what happened in the previous step.

Agents make sense when:

The task has clear success and failure conditions.
The cost of a wrong action is low or reversible.
You can afford for the agent to take longer than you would (it sometimes wanders).
The savings from "I don't have to be in the loop" outweighs the risk of errors compounding.

Agents are a bad fit when:

The task involves money, customer communication, irreversible changes, or anything where being subtly wrong is expensive.
You cannot specify success precisely enough for the agent to know when it is done.
The task requires judgment that depends on context the agent cannot see.

Coding agents (Claude Code, Cursor agent mode) are the most mature category. They work because compilers, tests, and linters give the agent fast feedback. The agent can try something, fail, try again, and converge on working code without you watching every step.

General-purpose "do this for me" agents are less mature. They sometimes get stuck, sometimes confidently complete the wrong task, and sometimes burn a lot of API tokens without producing useful output.

The rule of thumb: use an agent when you can verify the final result is correct without watching each step. If you have to watch each step, you should be using a workflow, not an agent.

The decision in one paragraph

If you do this task less than five times a month: prompt. If you do it often and want consistency: skill. If it has multiple steps that depend on each other or need human decisions between them: workflow. If it is repetitive, has clear success conditions, and the cost of a wrong autonomous action is low: agent.

Most people should be at prompts and skills 80% of the time, workflows for their highest-leverage recurring multi-step work, and agents only for tasks where the success condition is mechanically checkable.

How these compose

These are not exclusive. A workflow can use a skill at one of its steps. An agent can use prompts internally. A skill can wrap a prompt with consistent rules.

A practical example: a "weekly customer feedback synthesis" workflow might use a skill called "feedback-themer" at step 3 (cluster themes), a vanilla prompt at step 5 (draft the summary), and call an agent at step 6 (post the summary into Notion, Slack, and Linear). Each tool is doing what it is good at.

Common mistakes

A few traps people hit:

Building a skill or workflow for something they only do occasionally. Setup cost > savings. Use a prompt.
Trying to make a workflow autonomous (calling it an agent). If the steps need human review, do not skip the review.
Using an agent for customer-facing work without strong guardrails. Autonomous AI sending emails or making decisions for customers is where most agent disasters happen.
Skipping prompts entirely and starting at "let me build an agent." Most of what you wanted from the agent could have been a workflow you ran in 15 minutes.

Where each one lives on the site

Prompts: /prompts — 115 prompts across 10 categories (writing, code, sales, research, design, productivity, analysis, learning, creative, communications).
Skills: /skills — 132 hand-written Claude Skills across 12 role packs.
Workflows: /workflows — 164 step-by-step workflows across 5 packs (dev, sales, marketing, HR, founder). 40 free with email signup.
Agents: /agents — agent frameworks, building blocks, and what is mature enough to use today.

If you are deciding which to pick for a specific recurring task, start with the prompt version. If after a few weeks you find yourself re-using it constantly, graduate to a skill. If the work has dependencies between steps, build a workflow. Only consider an agent when the success condition is mechanically verifiable.