The CRAFT Framework: Build Reusable AI Skills

A 5-step way to build reusable AI helpers that get better over time, instead of typing the same prompt from scratch every day.

What is a reusable AI skill?

A skill is just a saved set of instructions you reuse, so you do not have to re-explain a task every time. You may have seen this as a Custom GPT in ChatGPT or a Project in Claude: you write the instructions once, and from then on the AI already knows the job. CRAFT is a simple way to build one of these and keep improving it. Each letter is a step, and none of it involves coding.

C: Capture What You Need AI to Do

Describe the task precisely. Vague goals make vague helpers.

A good skill starts with a clear job description. Not “help me with content,” but something like “take a news article, pull out the key facts, and write a short summary with a headline, a few sentences, and a link to the source.” The more precise you are, the more reliable the helper. Write down the whole picture: what goes in, what a great result looks like, the odd cases, and what it should never do. Think of it like writing instructions for a very literal assistant. Anything you leave fuzzy, it will guess at in its own way.

Example

A clear definition for a “newsletter quick take” helper: it takes a news article and produces an 80 to 120 word summary with a headline, a two or three sentence take, and a link to the source. It never speculates or adds opinion beyond the angle you give it. If the article is behind a paywall, it uses the headline and the preview text. If the story is more than three months old, it notes the date clearly. That is specific enough for AI to do it well every time.

R: Research Best Practices

Study how the best people do the task by hand. That is the standard you are aiming for.

Before you build a helper, look at how the best people already do the task. Building one to write LinkedIn posts? Study the top creators in your area. What patterns do they use? What makes their openers work? What is their usual shape? Collect the good examples, the templates, and the simple rules you notice. This is the step where you gather the raw material that will make your helper really good instead of generic, so do not rush it.

Example

Research for a “social media post” helper: save 20 posts that did well in your area, note the kinds of openers that get a response (questions, bold claims, short stories), jot down each platform’s basic rules (length limits, no hashtag spam), and keep a couple of bad posts as examples of what to avoid. Now you have everything the helper needs to learn from.

A: Arrange the Skill File

Lay out your instructions in clear sections so the AI follows them reliably.

A good set of instructions has a few clear parts, each doing a different job. Think of it like a short training manual for a new hire. Five sections cover it:

• The role: who the AI is playing, like “a newsletter editor writing for non-technical readers.”
• The steps: what to do, in order.
• The examples: a few of great work, and a couple of bad ones to avoid.
• The guard rails: the rules, like “never speculate” or “never use the word game-changing.”
• The output: exactly how the result should be laid out.

Use clear headings and keep each section short. The AI reads your instructions the way a new employee reads a manual, so clean structure helps it follow along.

F: Feed It Examples and Edge Cases

Show the AI what great looks like, and what bad looks like. Examples beat instructions.

Examples are the most powerful part of building a helper. One good example says more than a paragraph of instructions. Include two or three examples of an ideal result, and one or two of what to avoid, with a quick note on why each one is bad. Then think about the odd cases, which are where helpers tend to break: what if the input is really long, or really short, or in another language, or missing something? For each, spell out what to do. And keep adding over time. When the helper nails something, save it as a good example. When it gets one wrong, save the mistake and the fix as a bad example. That is how it gets better with use.

Example

A good example to give it might be a tidy summary: a clear headline, two or three plain sentences explaining what happened and why it matters, and a line crediting the source. A bad example to warn against: “This new gadget is a GAME-CHANGING must-have!!” with the note, “too much hype, all caps, exclamation marks, and no real substance.” Now the AI knows exactly which way to lean.

T: Test, Break, and Refine

Throw odd inputs at it. Find where it slips. A helper is not done until it handles surprises.

Testing a helper means deliberately trying to trip it up. Give it things you did not design for: a 10,000-word article when you built it for short ones, a piece in French, a broken link, a topic outside its lane. See what it does. Keep a simple log of what you put in, what you wanted, what you got, and whether it passed. When something fails, work out why and update the instructions. Did it need a new rule? A clearer step? An example for that odd case? Make the fix and test again. This is ongoing, not a one-time thing. After a handful of rounds, your helper will handle almost anything you send it.

Example

A simple way to test in rounds: first, try five normal inputs you expect it to handle. Second, try the odd ones: very short, very long, missing information, strange formatting. Third, try a few tricky inputs meant to confuse it, like a vague or contradictory topic. Fourth, re-run the normal ones to make sure your fixes did not break anything that worked before.

Key Takeaways

• Save your instructions: a reusable helper beats retyping the same prompt every day.
• Be precise: a clear job description is what makes a helper reliable.
• Lead with examples: showing good and bad work teaches faster than rules alone.
• Try to break it: odd inputs reveal what still needs fixing.
• It compounds: every example you add makes it better. See the full Framework System for the rest.

Frequently Asked Questions

Do I need technical skills to use this framework?

No. Every step is written for non-technical people. You write the instructions in plain English, and the tools we recommend all have free versions you can start with today.

Which AI tool should I use with this framework?

Any of them. It works with any tool, including ChatGPT (as a Custom GPT), Claude (as a Project), and Gemini. Start with whichever you are most comfortable with.

How long does it take to see results?

The first version helps right away. It really shines after a few rounds of adding examples and fixes, as your helper learns from real use and stops making the same mistakes.

Can I combine this with other Beginners in AI frameworks?

Yes, they are designed to work together. See our complete framework system for how STACK, BUILD, ADAPT, THINK, CRAFT, and CRON fit together.

Is this framework free?

The article and the method are completely free. We also offer the frameworks as downloadable PDF guides with extra templates and worksheets.

How do I build a reusable AI helper or Custom GPT?

Follow the CRAFT steps: capture exactly what you need it to do, research how the best people do it, arrange your instructions into clear sections (role, steps, examples, rules, output), feed it good and bad examples plus the odd cases, then test it by trying to break it and fixing what slips. Clear sections are what separate a reliable helper from a flaky one.

How do I create a Claude Project that improves over time?

The trick is the “Feed It Examples” step. Every time your Project produces something great, add it to the instructions as a good example. Every time it gets one wrong, add the mistake and the fix as a bad example. After a few rounds of this, your Project handles almost any request well, because it has learned from your real work, not just the original instructions.

The CRAFT Framework was developed by James Swierczewski at Beginners in AI. For more frameworks and practical AI guidance, subscribe to our free daily newsletter.

Sources

• Grokipedia: Prompt engineering
• Stanford HAI AI Index Report

You May Also Like

• The Best AI Prompts Library
• The STACK Framework: Better AI Prompts
• Best AI Tools for Beginners
• How to Edit AI Out of Your Writing
• The Full Framework System