Claude Skills explained: how progressive disclosure works, why the skill recorder beats manual writing, and how to build reusable skills fast.
What is a Claude Skill? #
A Claude Skill is a folder containing an instruction file that teaches Claude how to perform a specific, repeatable job. That’s the whole definition. Instead of re-explaining a process every time you chat with Claude, you package the steps once, give the folder a clear description, and Claude can call on it whenever a matching task comes up. Skills turn one-off prompting into a reusable capability, which is why they’ve become a core part of how people run real work through Claude instead of just chatting with it.
TL;DR #
- A Claude Skill is just a folder with an instruction file (a skill.md) plus any supporting reference files, built to handle one specific process reliably. - Skills rely on progressive disclosure, meaning only a short description loads into Claude’s context at first, and the full instructions and reference material load only when the skill actually activates. - The skill recorder lets you demonstrate a process directly inside Claude and have it generate the skill from that recording, which is faster and often more reliable than writing instructions manually. - Anthropic’s own skill creator skill gives you a consistent, modular template so skills stay reusable and can be chained together instead of built from scratch each time. - A clear activation description(when to use it, when not to, what it achieves) matters as much as the instructions themselves, because Claude has to find and trigger the right skill before any of that content helps. - Businesses getting the most value tend to adapt existing skills, including Anthropic’s prebuilt ones, rather than searching through large public skill marketplaces one by one. - Skills work best alongside static context files (brand voice, visual identity, business facts) referenced in a Claude.md, since skills handle process while those files handle recurring knowledge.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
How does progressive disclosure work? #
Progressive disclosure is the mechanism that makes skills efficient instead of bloating every conversation with unused instructions. When Claude scans available skills, it doesn’t load the full contents of each one into context. It only reads a short description for each skill, the summary of what the skill does and when it should trigger. Only once Claude decides a skill is relevant does it load the skill.md file itself, the step-by-step guide for that process.
From there, a second layer of disclosure kicks in. The skill.md typically references separate supporting files, additional context, templates, or examples that aren’t loaded automatically. Those only get pulled in if the specific step of the process actually needs them. So a complex skill might have dozens of pages of reference material sitting in the folder, but Claude only touches the parts relevant to the task at hand. This structure is why writing a good activation description matters more than people expect. If the description is vague or overlaps with another skill, Claude may never trigger it, or may trigger the wrong one. The description has to state clearly when the skill should activate, when it shouldn’t, and what outcome it’s built to produce.
How do you build a Claude Skill? #
There are two practical routes: write it manually, or record it.
Manual creation means building the skill folder structure yourself: a skill.md with numbered steps, a clear activation description, and any reference files the steps point to. Anthropic publishes its own skill creator skill, which gives a consistent, modular format to work from rather than inventing structure each time. Using that template matters if you plan to have several skills that reference each other or chain together, since consistency keeps them maintainable as the number of skills grows.
Recording is the faster route and a more recent capability. Instead of writing instructions, you perform the actual process inside Claude while it records your steps, then Claude generates the skill.md from that recording. This is particularly useful for complex, multi-app workflows where there’s no API to hook into and the only way to teach the process is to show it. A workflow with dozens of individual actions can be captured in a few minutes of recording, after which Claude can repeat it reliably without needing to be walked through it again.
The recording approach solves a real problem with manual skill writing: people are bad at fully describing their own processes from memory. A recorded demonstration captures the actual clicks, order of operations, and edge cases as they happen, rather than relying on someone to reconstruct them afterward.
Should you build skills from scratch or adapt existing ones? #
Other agents ship a demo. Remy ships an app. #
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Building every skill from a blank file is the slow path, and it’s largely unnecessary. Anthropic maintains its own library of prebuilt skills, including a plugin aimed at small businesses that bundles a set of skills (31 in total) covering common business processes. Adapting one of these to your specific workflow is usually faster than writing an equivalent skill yourself, and it starts you from a structure that already follows good practice around activation descriptions and progressive disclosure.
There are also public skill marketplaces, sites like skills.sh, where you can browse skills by popularity or trending status. These are worth knowing about, but sorting through a large public catalog to find something close enough to your use case takes real time. In practice, starting from Anthropic’s own prebuilt skills and adjusting them tends to get you to a working skill faster than searching a general marketplace, particularly for common business functions where Anthropic has already built something close to what you need.
How do skills fit into a broader Claude setup? #
Skills handle process: the repeatable “do this sequence of steps” work. They’re most effective when paired with static context that doesn’t change often, things like a brand voice profile, a visual identity reference, and a file of core business facts (ideal customer profile, product list, key links). That context typically lives in separate markdown files, with a central Claude.md file telling Claude which file to pull from for which kind of task.
The two systems complement each other. Static context files answer “what does this business sound like and look like.” Skills answer “how does this specific task get done, step by step.” A content-writing skill, for example, might reference the brand voice file as part of its instructions rather than duplicating that information inside the skill itself. Keeping these separate avoids repeating the same information across multiple skills and makes updates easier: change the brand voice file once, and every skill that references it stays current.
Frequently Asked Questions #
What’s the difference between a Claude Skill and a custom instruction in Claude.md?
Claude.md content loads into every conversation by default, so it’s suited to information Claude should always have on hand, like tone of voice or business facts. A skill only loads when it’s relevant to the task, using progressive disclosure, which makes it better suited to specific multi-step processes rather than always-on context.
Do I need to know how to code to build a Claude Skill?
No. Skills are folders of markdown instructions, not code. The skill recorder makes this even more accessible since it generates the skill.md from a demonstrated process rather than requiring you to write instructions by hand.
How does the skill recorder actually generate a skill?
You perform the task inside Claude while it records the sequence of actions and decisions you take. Claude then converts that recording into a structured skill.md with step-by-step instructions, which it can follow again on future requests without needing the process re-explained.
What is Anthropic’s small business plugin?
It’s a set of prebuilt Claude Skills, 31 in total, released by Anthropic and aimed at common small business workflows. It gives businesses a starting library to adapt rather than building every process skill from a blank file.
Can skills be chained together?
Yes. Building skills in a consistent, modular format, such as the structure produced by Anthropic’s skill creator skill, makes it possible for one skill to reference or trigger another, so multi-stage processes can be broken into separate reusable skills instead of one large monolithic file.