How to create more professional Claude skills with 8 content layers
Quick Summary
This article delves into the issue of AI "freewheeling" when using Claude Skills and presents a comprehensive solution: the 8-Layer Framework. This is a SKILL.md structure that helps clearly define the AI's role, context, inputs/outputs, rule set, and workflow. Additionally, the article introduces how to use Reference Files to help the AI learn real-world tone and style, ensuring consistent and user-aligned results.
You already know what skills are in Claude and have tried creating a few, but the results are still inconsistent. Sometimes the AI does exactly what you want, but other times it 'freewheels' in a completely different direction. The problem often isn't entirely with Claude; it lies in your SKILL.md file: missing layers, no order, or cramming everything into one long, unstructured paragraph.
The 8-Layer Framework is how the 4aivn team shares to address that, by breaking down skills into clear layers, from foundational to operational. Here, our team applies skills to deliver quality articles to you, and you can absolutely apply them to solve your own problems. But it needs to be said directly that this is only Prompt Engineering, just one part of a larger picture. For our team, this part helps us clearly assign tasks to the AI, but to produce truly high-quality articles that resonate with readers' psychology, follow conversion formulas, and meet our set standards, there's still much more to learn.
If you don't know what skills are in Claude yet, you can refer to this article: Claude Agent Skills are the skills to know about AI in 2026, which provides an introduction to approaching skills.
Why Skill Structure Determines Everything
SKILL.md, of course, looks quite similar to the long prompts many people have used with ChatGPT, Gemini, and Claude in 2024: "You are a copywriter expert with 10 years of experience, write articles using the PAS structure, empathetic tone, forbidden to use words like breakthrough, perfect solution..." After typing, chatting, closing, and opening a new chat next time, you have to instruct it all over again from the beginning.
Skills are different in that they are a guide you only need to write once, and Claude will understand how to work with you without needing re-explanation in each session. And now everything changes: a prompt defines what needs to be done this time, while a skill defines how to work in the long term. A common problem here is that many people write SKILL.md like a long prompt, cramming everything into one paragraph without layering. Claude can read it, but when it encounters a new situation beyond what you've listed, it lacks a conceptual framework to process it. That's why the results are often messy.
The 8-Layer Framework divides SKILL.md content into two groups: 4 foundational layers that help the AI know who it is and what it does, and 4 operational layers that help the AI know how it works.
Four Foundational Layers Defining Who the AI Is
Layer 1 – Mission
Define the core role of this skill. This is the first sentence Claude reads and uses to shape all subsequent behavior.
Example: "You are an editor specializing in writing and editing articles about AI for the 4aivn community, targeting Vietnamese people interested in AI but without a technical background."
Layer 2 – Context
Describe the environment in which this skill operates. The writing style for an "AI article" request differs completely between a website, a fanpage, or Instagram.
Example: "The article is published on 4aivn.com, readers primarily access it on mobile, requires short paragraphs, clear h2 h3 headings, and is approximately 1000 to 1200 words long."
Layer 3 – Input
Define what form Claude will receive information in. This layer is often overlooked, leading the AI to make assumptions when the input is unclear.
Example: "Input can be: a single keyword, a brief of a few lines, or a ready-made outline. If only a keyword is received, further questions must be asked before writing."
Layer 4 – Output
Define what the returned result should look like in terms of format, length, and default structure.
Example: "The default output is a complete article including an introduction (sapo), 3 to 4 H2 sections, and a conclusion. If the user only needs an outline, return a bulleted outline with a brief description of each section."
Four Operational Layers Defining How This AI Works
Layer 5 – Rule Set
This is the most important layer. You define the writing style, mandatory structure, and equally important, a list of things absolutely not to do. The more specific, the better.
Example:
- The introduction (sapo) must start with a real-world situation or a surprising statistic, never with a definition.
- At least 70% of H2 headings must be in question form to suit SEO, GEO, and each H2 must have at least one specific example.
- Forbidden phrases: "In the rapidly changing world of technology...", "It cannot be denied that...", "Hope the article is useful"
Layer 6 - Proactive Questions
Instead of the AI immediately diving into work, you make it ask questions first. This layer eliminates most cases of off-target results due to the AI making assumptions about the request.
Example: "Before writing any article, at least 3 questions must be asked: who is the target audience, what is the article's goal (inform/persuade/instruct), and what tone is desired (serious/approachable/neutral)."
Layer 7 - Plan
After asking enough questions, the AI must present an outline and clearly state the rules it will apply to this article before officially writing. You can see what it's thinking and can adjust its direction before it goes astray.
Example: "After receiving sufficient information, present: (1) a complete outline with a brief description of each section, (2) the main keywords and related keywords prioritized for this article."
Layer 8 – Agreement
Only when the user confirms agreement with the plan will the AI begin writing. Without this step, layers 6 and 7 are merely ceremonial, as the AI could still arbitrarily start after presenting the outline.
Example: "After presenting the outline, wait for user confirmation or requests for revisions. Only begin writing the complete article upon receiving a clear signal of agreement."
Write Your SKILL.md to Apply the 8 Layers Above
Of course, don't try to cram all 8 layers into one go. The practical order to start:
- Begin with Layer 1 and Layer 5 to clearly define the AI's role and rule set. Just these two layers will make the most significant difference compared to a regular prompt. Test it with 1 to 2 real-world requests to see if the output is correct.
- Once Layer 5 is stable, add Layer 6 to make the AI ask questions first. You'll immediately realize what information you often omit when assigning tasks, then add Layers 7 and 8 to close the control loop.
- Layers 2, 3, and 4 should be added when you notice the AI is making incorrect assumptions about the environment, input format, or output structure; these are signs that additions are needed.
References: A Very Important Part of Skills
After using skills for a while, you'll notice a new problem: the AI follows the correct structure and rules, but something about the brand is still missing, requiring further edits. The tone might be correct but not quite like yours, the structure might be right but not as familiar as your old articles.
This is where References come into play.
What are References in SKILL.md?
References are supplementary files you place alongside SKILL.md. They contain things that are too long or too specific to fit into the rule set, but which the AI needs to read in certain situations.
For content writers, the most useful type of Reference is approved output—the complete articles you've been satisfied with—used as reference samples so the AI can learn your actual tone and style instead of just reading abstract rules.
How Our Team Adds References to Skills
Folder structure:
writer-4aivn/
SKILL.md
references/
sample-article-01.md (published article, satisfactory result)
sample-article-02.md
sample-article-03.md
In SKILL.md, clearly declare when Claude needs to read them:
## Reference Files
references/sample-article-01.md: Read when the user requests a practical guide article
references/sample-article-02.md: Read when needing to reference the tone for an AI tool analysis article
Important Rule
Don't let Claude decide whether to read references on its own. Provide specific activation conditions, such as "read when the user requests an article of type X" instead of "read if needed." The latter is too vague, and Claude will either ignore it or read it at the wrong time.
How Many Sample Articles Are Enough?
Start with 2 to 3 sample articles of different content types: practical guides, tool analyses, opinion pieces. No more are needed in the initial stage. Each added sample article provides further evidence, helping the AI understand your tone one step better than just reading rules.
In summary, creating skills will take a lot of time initially, just like we spend a lot of time refining long prompts. But once stable results are achieved, we will often be surprised by what Claude can write and do.
This is the first part about article writing skills with AI. This initial part helps you get the first output from a skill, but the first output is rarely perfect. Subsequent parts will delve deeper into refining skills for more complex issues, until the AI works exactly as you intend.
