Blog
Claude projects vs skillshow to build Claude skillsClaude skill creator eval

Claude Projects vs Skills: What Goes Where and How to Test

Projects hold who you are; skills hold how a job is done. Learn what goes where in Claude, why connected files beat uploads, and how to test a skill.

Portrait of Luis Chavez-Mattos
Edited by Luis Chavez-Mattos, Director, Product··RSS

For Operations, Managers and Engineering

A freelance copywriter in her late 20s slides one card into an open project folder, with two more folders open beside it ready for the same card.

What is the difference between a Claude project and a skill?

A Claude project holds the context about who you are: your files, brand voice, clients and background. A skill holds how one specific job gets done: the steps, standards and fixes that make the output right. A simple rule decides where anything goes. If it describes who or what, put it in the project. If it describes how a particular job is done, put it in a skill. Skills can then be reused across many projects.

TL;DR

  • Projects store context once, so every chat inside them starts already knowing your files, voice and data sources.
  • Skills store the method for one job, and they work across multiple projects instead of being copied into each.
  • Use one project per set of context, not one per task. If two jobs need the same files, they can share a project.
  • Prefer connected files over uploaded ones. An upload is a copy that goes stale. A connected Google Doc is read fresh every time.
  • Build a skill from a list of your own corrections, not from a blank page, and keep only the fixes Claude can't get right on its own.
  • Test the skill against no skill using the eval inside Anthropic's skill creator skill. An untested skill may be doing nothing at all.

What belongs in a Claude project?

A plain chat has no reusable context. Every new conversation, you retype who you are, what you sell and how you write. A project fixes that. You put information in once, and every chat inside the project starts with it.

Think of a project as inputs and outputs:

  • Inputs: your files, instructions, brand voice, visual identity, and live data from apps you connect through MCP (a standard way to plug outside tools into Claude) or APIs.
  • Outputs: drafts, files and even live dashboards, which Claude produces as artifacts.

Here is the example the source uses. Say you want to get cited in AI search answers. Your project holds your business files and your writing voice, and it reads search data live from a tool like Ahrefs. Because the context is already there, your weekly prompt can shrink to one line: "Find this week's opportunities." Claude reads that as an SEO request and looks for pages losing traffic and questions where ChatGPT cites a competitor instead of you.

That is the payoff. The prompt no longer has to explain the world. It only states the goal.

How many projects should I make?

Make one project per set of context, not one per task. The test is a single question: would these two jobs need the same files?

  • If yes, they can live in the same project.
  • If the voice or the client changes, so the files change, split them into separate projects.

A freelance copywriter, for instance, would keep a separate copywriting project for each business, because each has its own voice and materials. A marketer at one company writing emails, landing pages and social posts in a single brand voice can keep all of that in one project.

Why use connected files instead of uploads?

An uploaded file is a copy. The moment you edit the original, the copy is out of date, and Claude keeps working from the old version.

A connected file, such as a Google Doc linked through MCP, is read fresh each time. Update your voice brief once and every project that points to it gets the new version.

This also helps teams. Share the Google Doc with the right permissions, and everyone using it in a conversation gets the latest document. Nobody has to re-upload anything or ask which version is current.

Practical habit: anything you expect to edit more than once (voice guide, pricing notes, client brief, product positioning) should be connected. Anything that never changes (a signed contract, a published report) is fine to upload.

What belongs in a skill?

A project should hold broad context. It should not hold how you like one specific job done. That belongs in a skill.

The old advice treated skills like instruction manuals for a task. A better way to think of it: a skill should contain only the fixes Claude can't get right out of the box. Claude already knows how to write an email or summarize a report. What it doesn't know is the particular ways your version of that job goes wrong.

This matters for another reason. Newer Claude models tend to do more than you ask rather than less, and, as the source puts it, prescribing rigid step-by-step instructions can limit how the model works. Skills that spell out individual steps can have the opposite effect, limiting the model to one way of working. Skills that capture specific corrections and standards work better.

How do I build a skill from a corrections list?

Don't start from scratch. Do the job first, correct the result, and let the corrections become the skill.

  1. Run the job without a skill. Use the same prompt you would normally use.
  2. Correct the output until it is what you wanted.
  3. Ask Claude to keep a list of every correction you made.
  4. Sort the list into three piles.
  5. Add examples and let Claude draft the rules.
  6. Test it (covered below).

Here is a prompt you can copy for steps 3 and 4:

For the rest of this conversation, keep a running list of every
correction I make to your output. When I say "sort the list",
group the corrections into three categories:

1. About me or my business (belongs in project context)
2. General AI writing habits that are wrong in any job
   (for example, generic hooks or filler phrases)
3. Specific to this job (belongs in the skill)

For each item, note what you did and what I changed it to.

Where each pile goes

Type of correctionWhere it goesExample
About you or your businessThe projectYour audience, tone, offers
General AI habitsAn existing skill, if one covers itA humanizing skill such as Stop Slop for "it's not X, it's Y" openers
Specific to this jobYour new skillRules for how a page must be structured to get cited

The middle pile is worth noticing. If someone has already built a skill that fixes generic AI phrasing, use it. Don't rebuild that inside your own skill.

Add examples

Once you have the job-specific rules, layer on examples. For the AI search case, that might be two of your pages that do get cited in AI answers. Claude can work out why those pages worked and turn that into rules. Examples give the skill a standard to aim at, not only a list of things to avoid.

How do I know a skill actually works?

Test it against doing the job with no skill. This sounds obvious, and most people skip it.

Run the same job with the same starting prompt twice, once with the skill and once without. Anthropic's skill creator skill includes an eval function that compares the two like for like against a set of checks. You see how each version performed, not a vague impression of which felt better.

Why bother? Because a skill you haven't tested might be doing nothing. Claude may already handle the job well enough that your skill changes little. Or the skill may add rules that don't improve the checks you care about. The comparison tells you which.

A sensible way to set it up:

  • Write the checks before you run the test. For example: "every claim links to a real source" or "the page follows our heading structure."
  • Use the same prompt for both runs.
  • Keep what moves the results. Cut rules that make no difference.

The same idea applies to ordinary prompts. The source recommends giving Claude a way to check its own work, such as requiring that every question it surfaces links to a real result it can prove. A good prompt states the goal and context, what done looks like, how to validate, what is out of scope, and what it can't touch.

What happens when I want the job to run without me?

Once the project holds your context and the skill holds your method, you can hand the job off. The source describes a scheduled task: a short brief made of your original prompt, the skill it should use, where the output goes, and what it must not touch. For AI search, that could be four lines: find the top opportunity from this week's data, draft the page with the skill, save it as a draft in a folder, and don't publish anything.

Tasks can start on a schedule (every Monday at 9 a.m.) or from an event, such as a new pull request or an API call from your own tool. By default they run in the cloud, so your laptop can be closed.

Keep a human in the loop based on risk. The creator in the source never hands over anything involving money, complaints, legal matters, or anything going out under their name for the first time. That is a reasonable starting list for anyone.

What this means for you

If you use Claude at work and keep repeating yourself, you have two separate fixes.

  • Repeating who you are or what your company does? Build a project and connect your key documents instead of uploading them.
  • Repeating the same corrections to the same kind of output? Start a corrections list on your next task, then turn it into a skill.
  • Managing several clients or brands? Separate projects for separate context, but share one skill across all of them, so a fix made once helps every client.
  • Skeptical your skill helps? Run the with and without comparison before you trust it.

Start small. Pick one job you do weekly, do it once, and keep the list of corrections. That list is most of the work.

Can I build a tool to track my corrections and skill tests?

Once you are running this method across several jobs, the corrections lists and test results tend to end up in scattered documents and spreadsheets. A small internal app could hold them in one place: a form where you or your team log each correction, tag it as project context, general habit or job-specific, and record how each skill scored with and without it. With Remy, you describe that tracker in plain English, and it builds and deploys it as a working web app with a database and sign-in, so you own it instead of adding another subscription.

Frequently Asked Questions

5 questions

Should my instructions go in a Claude project or a skill?

Anything about who you are or what your business is goes in the project. Anything about how a specific job should be done goes in a skill. Skills can be reused across projects, which is why the how-to material is kept separate.

Is it better to connect files or upload them to a project?

Connect them when you expect the file to change. An upload is a copy that goes out of date as soon as you edit the original. A connected Google Doc is read fresh each time, and a shared one gives your whole team the latest version.

How do I test a Claude skill?

Run the same job with the same prompt, once with the skill and once without. Use the eval function inside Anthropic's skill creator skill to compare the two against a set of checks. If the results don't differ, the skill isn't earning its place.

Do I need a new project for every task?

No. Use one project per set of context. If two jobs need the same files, they can share a project. Split projects only when the voice, client or files change.

Should I write a skill from scratch?

Usually not. Do the job once without a skill, correct the output, and keep a list of your corrections. Sort that list, and build the skill only from the fixes specific to your job.

How we source, write, and correct our articles. Editorial standards

Related articles

Straight to your inbox.
New pieces as they publish. No noise.