The 8 Claude Skills That Changed My Work (and One to Avoid)

The 8 Claude skills I actually use, updated October 2026: from claude-md-management to video-use. The one to avoid is among the most installed in the world. And the skills my clients use, department by department.

Gaetano Castaldo Gaetano Castaldo
13 Mar 2026
18 min read
Updated 12 Oct 2026
Most read Claude #claude-skills #workflow #ai-tools
The 8 Claude Skills That Changed My Work (and One to Avoid)

Updated October 2026. The Claude skills I actually use are eight: claude-md-management, claude-automation-recommender, mcp-builder, skill-creator, Anthropic's document skills, n8n-skills, video-use and process-builder. The one to avoid is frontend-design: it is among the most installed skills in the world, and I removed it.

# Skill What it does Published by
1 claude-md-management Keeps CLAUDE.md, the project memory, up to date Anthropic, official plugin
2 claude-automation-recommender Tells you which skills, hooks and agents your project needs Anthropic, claude-code-setup plugin
3 mcp-builder Guides you through building an MCP server Anthropic
4 skill-creator Creates and tests new skills Anthropic
5 docx, xlsx, pptx, pdf Produce real Office files and PDFs Anthropic
6 n8n-skills Makes Claude write correct n8n workflows Community
7 video-use Edits videos by talking to Claude Community, Browser Use
8 process-builder Turns a process interview into a BPMN diagram Castaldo Solutions, open source
To avoid frontend-design Web interface design Anthropic

When Anthropic introduced skills in October 2025, few understood right away what was happening. It was not a feature: it was a protocol. Today it is an open standard, Agent Skills: Anthropic created it, and the official showcase lists more than forty products that adopt it, including ChatGPT and Codex, GitHub Copilot, Cursor and Gemini CLI (checked on 12 October 2026). Same path as MCP: born at Anthropic, now the industry standard.

Skills today are what plugins were to browsers: those who master them work in a structurally different way from those who don't yet. In this article I tell you about the eight I use, the one I removed, and what my clients actually use.

What is a Claude skill (and why is it not just a prompt)?

A skill is a folder with a SKILL.md file at its core: instructions, references, examples, sometimes scripts. Claude activates it when the task matches its description, or you call it yourself with a command such as /deploy.

The difference from a long prompt in chat:

  • It is modular. At startup Claude reads only each skill's name and description, about 100 tokens according to the Agent Skills specification; the rest comes in only when needed.
  • It is shareable. It lives in the repository, goes into git, and the whole team uses it identically.
  • It is versionable. Change the skill and you change the behaviour, without touching the code.

Can a skill make Claude worse?

Yes, and you need to know it before choosing which ones to use.

On 14 July 2025 Chroma published Context Rot, a study of 18 models (including Claude Opus 4, GPT-4.1 and Gemini 2.5 Pro): across all experiments, performance degrades as the input grows. The thresholds differ from model to model: in the repeated words experiment, Claude Opus 4 and GPT-4.1 start refusing at around 2,500 words, and GPT-4.1 nano starts writing a lowercase "san" where it should have copied "San Francisco". What it means for anyone working in a chat, with per-model charts and real sessions, is in the article on context rot.

The practical point: a skill with too heavy a payload can make the model worse instead of better. You do not pay the weight at startup, you pay it when the skill activates and its whole SKILL.md enters the context. The specification recommends staying under 5,000 tokens; my rule, which is mine and not the specification's, is stricter, and you will see it applied to frontend-design below. When there is a lot of documentation to bring in, an MCP works better, because it brings it on demand instead of loading it all up front.


What are the best Claude skills? The 8 I actually use

1. claude-md-management: The Most Valuable of All

This is the one that changes everything for teams.

claude-md-management manages the CLAUDE.md file: the "project memory" Claude reads at every session. It contains the project rules, conventions, critical files, deploy workflow, brand. Everything a new dev (or a new AI session) needs to know before touching the code.

The revolutionary point: it lives in the repository. It goes into git. Every developer who clones the project has the same shared memory. You are not configuring Claude: you are documenting the project in a format AI reads natively.

The skill has two commands:

  • /revise-claude-md: updates CLAUDE.md with the learnings from the current session
  • claude-md-improver: audits and improves the quality of the existing file

I use this system on all Castaldo Solutions projects. The result: whoever joins a project starts from the same rules as those who were there before, and no session starts from scratch.

2. claude-automation-recommender: The Skill to Optimize Skills

Paradoxically, one of the most powerful skills is the one that tells you which other skills to create.

claude-automation-recommender analyzes your codebase and workflow, then returns a concrete map: which agents to create, which hooks to configure, which plugins to activate, how to structure local skills. It tells you exactly how to optimize your Claude setup for the specific project.

I used it to build the deploy automation of this site: workflow analysis → identification of repetitive patterns → custom /deploy skill → automatic post-deploy smoke test. A loop that now completes in a few commands.

It is the skill that takes you from "I use Claude in chat" to "I have an AI infrastructure on the project".

3. mcp-builder: Creating Your Local MCP (With Caution)

This skill guides you through creating a local MCP server: your custom application reachable from Claude as a native tool.

I have applied it on several projects: connections to proprietary databases, integration with internal APIs, access to legacy systems. When it works, it is magic: Claude talks directly to your systems without copying and pasting data.

The trap, though, is concrete: many use cases are already covered by the community. Before building an MCP from scratch, search GitHub and the official marketplace: there is a high chance someone has already done it better. The risk is losing days replicating something that already exists. mcp-builder is powerful when your use case is genuinely custom; it becomes a waste when you are reinventing the wheel.

4. skill-creator: Writing the Skills That Don't Exist

Sooner or later the skill you need does not exist, because it describes the way you work, not someone else's. skill-creator is Anthropic's official skill for building it: it guides you with questions, generates the folder structure and the SKILL.md, and helps you test it.

I add two rules I give all my clients:

  • First, the empty test. Do the task without any skill and note where Claude really goes wrong. The skill should cover those mistakes, not repeat what the model already knows. It is the fastest way to keep it short.
  • The description is the trigger. It is the text Claude reads to decide whether to activate the skill, and the specification caps it at 1,024 characters. Two skills with similar descriptions compete for the same task, and Claude chooses badly.

If you want to build one step by step, I wrote a practical guide with real examples; to maintain them over time, the best practices I use.

5. Anthropic's document skills: docx, xlsx, pptx, pdf

They are the least spectacular and the most used in my work with clients. They make Claude produce a real Word file, spreadsheet, presentation or PDF, with styles, tables and formulas, instead of text you have to copy by hand.

Document skills matter because a client reads a document in Word or PDF, not in markdown. One practical lesson: when client-facing text ends up inside a script that generates the file, check the accents. The first Excel file delivered this way had them wrong, and since then it is a fixed check before every delivery.

6. n8n-skills: Specialized Automation

n8n-skills is the perfect example of a vertical skill for a specific tool: a collection of 15 skills (as of 12 October 2026) for building n8n workflows correctly.

n8n-code-javascript, n8n-expression-syntax, n8n-workflow-patterns, n8n-validation-expert: each brings specific, structured context for a different aspect of n8n. The model knows how to use $input, $json, $node, how to handle errors, which patterns to apply for webhooks, databases, AI agents.

The value: without these skills, Claude produces n8n code that looks plausible but is often wrong: incorrect syntax, non-idiomatic patterns. In my experience, with the skills the first attempt holds up much more often. A vertical skill beats a generic prompt.

7. video-use: The Video Pipeline Without an Editor

It is one of the skills I use most right now, because it automated my marketing pipeline. video-use is open source, published by Browser Use, and has almost 29,000 stars on GitHub (12 October 2026).

It works in a clever way: instead of making the model watch the video, it turns it into text, a word-by-word transcript, and on that Claude decides cuts, subtitles, color correction and on-screen graphics. You talk, the video gets edited. You need ffmpeg and an ElevenLabs key for the transcription.

I use it to go from the raw recording to a video ready for LinkedIn and YouTube: a job I used to outsource or put off, and now I finish it in session. I also installed it for a client, to edit reels with subtitles.

8. process-builder: From a Process Story to a Diagram

The last one is mine, and it is the only one of mine on this list because it is public. process-builder turns an interview about a business process into a BPMN swimlane diagram, readable by a person and by an agent: the current version (AS-IS) with its pain points, and the future one (TO-BE) with the first fixes and full automation.

I released it as open source, under the MIT license, because the value is not in keeping it to myself: the more consultants use it, the more the way of describing a process becomes a shared standard, as already happened with MCP. You will find it with the other public tools in the Open Lab.

The skills I built for my own work are not on this list: they fit the way I work, not the way you do. Compared with the first version of this article I removed ralph-loop, which I no longer use, and brand-guidelines, which applies Anthropic's brand, not yours; n8n-skills used to be called n8n-mcp-skills.


Which Claude skill should you avoid?

frontend-design, Anthropic's design skill. And it is a surprise, because it is among the most installed in the world: eighth in the skills.sh ranking, with about 979,000 installs (all-time total, read on 12 October 2026).

On paper it looks perfect: it guides Claude in creating high-quality web interfaces, distinctive components, production-grade code.

The problem is the weight. I measured it with claude plugin details in September 2026: about 2,700 tokens every time it activates, as I explain in the article on how to install skills. And here context rot comes back: in my experience the model was spending attention on instructions that, in the meantime, it had learned to follow on its own. In March I considered it a skill to use with judgment; in September I uninstalled it completely.

In its place I use an MCP like context7, which brings the official documentation of frameworks and libraries on demand, only when needed.

The lesson: not every skill fits every use case. My rule of thumb: if a skill brings more than 2,000 tokens of instructions to a creative or exploratory task, consider an on-demand MCP instead of a static skill.

Which skills do companies use?

The companies I work with mostly use skills built on their own procedures, department by department. To write this paragraph I went back through four months of my meeting notes and my emails with clients: the only public skill they use is video-use, all the others are custom-built.

Department by department:

  • Administration. At an engineering firm, the administration team processes the month's attendance and meal vouchers with a skill. In the first month a defect showed up, merged cells in the Excel file, fixed in the next version: a skill gets tuned on real use, not on the demo. At a food company, the payroll journal entry is in testing: it starts from the payroll provider's summary, builds the entry with the right accounts and checks that it balances.
  • Production and quality. At the same food company a skill checks the packaging sequence against allergens, and production uses it directly.
  • Financial advisory. Skills that analyze case dossiers and retrieve company records, always asking for confirmation before buying them.
  • Marketing. video-use to edit reels.

The rules I give everyone, beyond the empty test and the description:

  1. After every update, a new conversation. In the old one Claude still has the previous version's instructions.
  2. Files in the folder, not pasted into chat. The skill reads them from there, and the conversation stays light.
  3. Every installed skill has a cost, even when you don't use it: how many to keep is a decision, not a detail.

What are the most installed skills in the world?

The ranking is dominated by skills for developers:

Skill or collection What it measures Value
find-skills (Vercel) installs, first in the ranking 3.8 million
grill-me (Matt Pocock) installs 1.3 million
frontend-design (Anthropic) installs, eighth 979,000
hyperframes (HeyGen, video) installs, five entries in the top 25 up to 959,000
superpowers (obra) GitHub stars 298,000
anthropics/skills GitHub stars 181,000
video-use (Browser Use) GitHub stars 28,800

Sources: the skills.sh ranking, all-time total, and the projects' GitHub pages, both read on 12 October 2026.

Two takeaways from this table. Video is the most represented non-code category in the ranking: HeyGen has five skills in the top 25, and it is the same need that put video-use on my list. And the most installed are not necessarily the best for you: the third in this table is the one I told you to avoid.

Claude Has Entered My Consulting Pipeline

The biggest change of these months is not a single skill: it is where Claude has slipped into the way I work with clients.

I record the meeting and transcribe it with Gemini. From there Claude digests the transcript: it extracts decisions, constraints, numbers, what the client actually said, not what I remember the evening after. That material becomes the scaffolding, the starting structure, of the company brain and of the projects: a ready structure, not scattered notes to reorder in hindsight. From there, in some cases, we get to agentic delivery: Claude directly executing part of the agreed work, not just preparing it for me.

My thesis: agentic delivery in consulting is not yet an established model. Few really do it, even fewer do it well. But it is gold, because it is the difference between selling hours and selling a result that arrives sooner and costs less to produce.

A note on the move to the Claude 5 family: when Fable 5 arrived I rewrote nothing. Skills, CLAUDE.md and MCPs kept working identically on the new model. The investment is not in the prompt, it is in the infrastructure, and the infrastructure survives a change of model.

If there is a thread tying all of this together, it is a simple idea: a skill is not just any automation. An automation runs a fixed step, always the same, without looking around. A skill virtualizes an assistant that thinks and acts as a team member would: it reads the context, decides what needs doing, does it. It is the difference between a script and a colleague you explained the job to well, once.


Why did skills become a standard?

Because the format works on any agent: a folder, a SKILL.md file, a description that says when to use it. Agent Skills made it an open standard, and today the same skill runs on Claude, on ChatGPT and Codex, on GitHub Copilot and on dozens of other products. As happened with MCP, we are watching a shared protocol emerge for defining what AI agents can do.

The implications for anyone working with AI today:

  1. Skills are investments, not configurations. A well-built skill lasts for months, works on every agent that adopts the standard, and can be shared with the team.
  2. Context rot is real and measurable. Before adding a skill, ask yourself: does this extra context help or distract the model?
  3. The real value is in the whole. A single skill is useful. A system of skills, hooks, CLAUDE.md and MCPs on a project is an AI infrastructure. The full picture is in the guide to Claude Code.

The same MCP protocol is what lets Claude talk to your business systems: if you have a CRM, you can connect it to an AI agent with the practical guide to MCP servers for CRM.

If you are building something serious with Claude (a product, a business workflow, a consulting system), the starting point is not "which prompt do I write" but "which infrastructure do I build".

Frequently asked questions

What are the best Claude skills?

It depends on what you do, but the ones I actually use in October 2026 are eight: claude-md-management for project memory, claude-automation-recommender, mcp-builder, skill-creator, Anthropic's document skills (docx, xlsx, pptx, pdf), n8n-skills, video-use for video and process-builder for processes. The criterion is the ratio between what they do and the context they consume: a heavy skill on a task the model can already do can make the result worse.

Are Claude skills the same as "competenze"?

Yes: on Anthropic's Italian pages skills are called "competenze". A skill is a folder with a SKILL.md file containing instructions, examples and sometimes scripts for a specific task. At startup Claude reads only each one's name and description, and loads the rest when the task matches. They work in Claude, in Claude Code and through the API, and in Claude Code you install them as folders or through plugins.

How many skills should you install on Claude?

The ones you use, and no more. Every installed skill takes up context even when you don't use it, because Claude always reads its name and description, and when it activates it loads all its content. Several skills with similar descriptions compete for the same task and Claude chooses worse. The criterion I use to decide what to keep and what to remove is in the article on how to install Claude skills.

Do Claude skills work with ChatGPT too?

Yes. The format became an open standard, Agent Skills, created by Anthropic and adopted by more than forty products: among them ChatGPT and Codex, GitHub Copilot, Cursor and Gemini CLI. A well-written skill, with a folder, a SKILL.md and a clear description, generally works on all of them without rewriting: what changes is where you install it and the commands to call it.

Who is Castaldo Solutions and what does it do with Claude in companies?

Castaldo Solutions is a consulting firm with its registered office in Milan and its operating office in Pavia, bringing AI and automation to Italian SMEs with up to 50 employees. With Claude it builds skills and agents for the tasks departments repeat every day, and writes the usage rules when the tool has already spread through the company without governance. The method is the CompanyTech BattlePlan, four phases over 4 to 8 weeks, and the first step is a free Pre-Assessment.

Tags

#claude-skills #workflow #ai-tools
Gaetano Castaldo
Gaetano Castaldo

Founder & CEO · Castaldo Solutions

Sono un consulente di trasformazione digitale con esperienza enterprise. Aiuto le PMI italiane ad adottare AI, CRM e architetture IT con risultati misurabili in 90 giorni.

The tools you just read about are public

Skills, MCP servers and the projects we use every day live in the Open Lab, with the code in plain sight.

Go to the Open Lab

Open code, no sign-up

Read also

Related articles you might find interesting