Practical guide to Claude Skills, projects, and Claude Code: how to build your AI stack

Claude Skills (Agent Skills) is a way to package your experiences and processes once so that the model can reliably repeat them without endless explanations in each new chat. In this article, we will explore what Skills are, how they are structured, how they differ from projects and MCP, and most importantly - how to apply this approach in tasks for channel authors, advertising buyers, SMM specialists, and video creators.

What are Claude Skills

Claude Skills are modular capsules of knowledge and processes for Claude:

• You describe once how exactly to perform the task (format, criteria, templates)

• You save this as a skill

• Then Claude understands from the context when this skill is applicable and loads it

So instead of writing every time: Here is our document/post/report format, here is how we measure metrics, here is our tone - you format this into a skill once, and then the AI remembers and reproduces it.

In general, a Skill is a folder with structured content that Claude can read and apply:

• Markdown files with instructions and an example of the result

• Auxiliary files (templates, examples)

• If needed, executable code (scripts for data analysis, file handling, etc.)

• Metadata: when the skill is relevant, what commands/formulations activate it.

Claude dynamically loads only those skills that are needed for a specific task to avoid cluttering the context. By the way, you might find my roadmap for studying Claude Code useful; it contains free training, helpful materials, guides, collections of MCPs, sub-agents, and commands.

Why Skills are now emerging in almost all discussions about agents

A bit of history to understand why Skills are now emerging in almost all discussions about agents:

• October 2024
  Anthropic quietly launches Agent Skills in beta: access via API and Claude Code. The target audience is developers who want to describe repetitive processes as separate capability packages.

• December 18, 2024
  Skills appear in the web interface of claude.ai, organizational-level skill management (Team / Enterprise) is added, a skills catalog from partners (for example, from Notion, Figma, Canva, Atlassian) is released, and an open standard for Agent Skills is published at agentskills.io.

• 2025
  Skills begin to connect with other capabilities of Claude: memory, augmented thinking, and the Cowork agent. Interfaces for building skills without code emerge.

• What’s happening now (February 2026)
  Skills are available in the Pro, Max, Team, and Enterprise plans. You can use ready-made skills (Excel, PowerPoint, PDF, etc.); implement partner skills from the catalog; or create your own tailored to your processes.

The strategy of Anthropic is clear: first - developers, testing the idea, and only then - user-friendly interfaces for regular users.

Why Skills are needed by office workers, authors, marketers, and everyone in general

In any job, whether it's a channel author, SMM specialist, advertising buyer, video editor, or project manager, the same things regularly repeat:

• Documentation only according to company standards, onboarding new employees, preparing reports according to regulations, analyzing operational meetings, and other corporate routines

• If you are an author, you need to write in a specific tone, consider different audience levels, account for various post orientations, avoid certain slang and FOMO, and so on

• An advertising buyer needs to consider certain criteria each time when choosing advertising channels

• Video authors need to structure the script of the video according to a specific format each time

• If you are launching a product, you need to provide the LLM with a warm-up template and explain the launch logic each time.

In general, you understand that regardless of the responsibilities (of course, if it’s not a physical task or other limitations, which I will discuss below), without skills you explain this to the AI from scratch each time. With skills (Skills), you describe it once, onboard it, so to speak, and then you work with it as with a trained assistant.

How Skills Work

To put it simply, the mechanism is as follows:

1. Finding the right skill. When you set a task, Claude looks at the list of available skills and determines which ones are relevant for that request.

2. Gradual loading. Only what is needed is loaded: not your entire package of instructions and processes, but specific skills that are involved in the current request.

3. Automatic invocation. You do not need to write use skill "X". Claude understands from the phrasing that it's about preparing an order / analyzing channels / scripting a video, and pulls in the necessary skill.

4. Combination of skills. You can use several skills simultaneously: for example, “channel style” + “news template” + “audience reaction analysis”.

5. Executable code (optional). In Skills, there can be scripts: then Claude not only knows the procedure but actually runs the code, for instance, to parse CSV by advertising channels.

Claude Skills + Claude Code: How It Looks in Live Work

In the Claude interface, there are two closely related mechanisms:

• Agent Skills - a formal standard of skills (a folder with manifest.json, instructions.md, etc.), which can be used via API, Claude Code, and organizational settings.

• Claude Code Projects - folders on your disk where Claude Code sees files, reads them, and acts according to what you have described in them.

In practice, the approach is the same:

You turn your head and work experience into a set of files: instructions, examples, criteria, and Claude acts as an assistant that strictly follows all of this.

Next, I will provide specific cases from my daily work: channel, advertising, video, launches.

Case 1. Content plan and posts for a Telegram channel

The channel needs to be maintained regularly: news, service analyses, reactions to information triggers. Every time, you have to:

• rebuild the delivery

• monitor the tone

• clean up excessive geek jargon

• develop a hook, introduction, and conclusion separately.

This is not difficult, but it constantly consumes attention.

How to package this into skills / project Claude Code

In the Project, I keep several files:

• CHANNEL_STYLE.md - channel tone, length, stop words, examples of successful and unsuccessful formulations.

• POST_TEMPLATES.md - types of posts (news, service analysis, response to information prompts, explanation).

• AUDIENCE.md - who reads the channel, what pains they have, what level of immersion in the topic.

Claude Code:

• reads these files

• does not forget the style from session to session

• does not require reminders like write in a human way, without unnecessary bureaucratese each time

My typical task:

Take this news, make a post for Telegram in the format of a calm analysis: without hype, with examples of application, without complex terminology. Add a strong introduction and a neat conclusion, without direct sales.

What I get at the output:

• Posts look consistent in tone

• Fewer revisions

• In one session, you can prepare several posts at once, instead of squeezing out just one.

In this mode, Claude is no longer just a text generator, but a full-fledged editor who knows your channel almost as well as you do. After all, you have loaded the knowledge and context, how it should be and how it shouldn't be.

Case 2. Purchasing ads in other Telegram channels

Advertising in Telegram is:

• analysis of channels (reach, ER, growth/fall, inflations)

• comparison of price per thousand impressions/subscribers

• tracking past placements and their effectiveness

• correspondence with admins.

Ideally, everyone should maintain a procurement table that reflects prices, agreements with admins, advertising creatives, and other necessary information. But usually, everything is scattered across tables, screenshots, and personal notes.

You have read to the middle of the article, which means that the article probably appealed to you, as did my presentation and writing style, so you can support me on my Telegram channel, where a lot of useful and verified information is already gathered.

How to package this

The project includes:

• CSV / Excel with channel statistics

• screenshots of reach and examples of placements;

• history of past purchases;

• AD_CRITERIA.md - my criteria and red flags (suspicious spikes, strange ER, discrepancies between declared and actual reach).

Claude Code:

• reads tables

• matches data between channels

• finds anomalies

• forms meaningful textual conclusions.

My typical task

Analyze these 15 channels, rank them by effectiveness for purchasing, highlight the top 3 and 3 questionable ones. Explain for each why you think so, based on the criteria from AD_CRITERIA.md.

What to expect

• Not just a dry table, but a clear report that identifies risks and potential.

• Fewer impulsive, ill-considered decisions, more reliance on criteria.

• Easy to explain to a partner or client why you are choosing these channels.

Here Skills/project already plays the role of an advertising analyst.

Case 3. SMM and short videos (Reels / Shorts / TikTok)

Short videos require:

• an idea and a clear hook in the first seconds

• a script

• on-screen text

• voiceover lines

• adaptation for different platforms.

If you do this from scratch every time, production quickly hits a ceiling in terms of time.

Structure of the skill / project

• VIDEO_FORMATS.md - video formats (explanation, reaction, myth breakdown, mini-tutorial, etc.).

• HOOKS.md - a set of hooks that really worked with a brief description of where they are appropriate.

• PLATFORMS.md - differences in formats and limitations of Reels, Shorts, TikTok.

Claude Code:

• selects the format and hook for the task

• writes the script and on-screen text

• adapts it for the chosen platform.

My standard task:

Create a 40-second video script: calmly explain service "X" for authors, without shouting or info-gypsy tactics. Add a hook in the first 3 seconds and brief on-screen text that duplicates key thoughts.

As a result, I receive scripts that can be read directly to the camera. Additionally, I get familiar formats that make it easier to shoot a series of videos. There is a sense of system rather than chaotic content.

Case 4. Editing and Voiceover: How to Speed Up Production

Even if a separate person does the editing, they still need:

• technical specifications

• clear timings

• text for voiceover

• notes ("pause here", "insert screenshot here", "close-up here").

When all this information is passed verbally or in writing, the number of revisions increases.

Solution with Skills /Claude Code

I simply throw in the draft text or transcript, describing the format of the video (review, tutorial, case).

Claude:

• breaks the text into fragments by timecodes

• creates a version for voiceover: short phrases, convenient for breathing

• adds notes for editing: where the pause is, where to insert graphics, where the speaker's shot is larger.

Example Task

Adapt this text for voiceover for a YouTube video: short phrases, without long constructions, indicate places for pauses and visual inserts.

As a result, the voiceover turns out to be livelier and calmer, it’s easier for the editor to navigate, and the project bounces less between versions.

Case 5. Launches, Special Projects, and Partnerships

Launching a product or worse - a large special project - includes:

• warm-up (a series of posts and videos)

• main presentation of meaning

• working with objections

• advertising creatives and landing pages.

If each element lives on its own, it results in a disjointed picture and a sense of chaos.

Skill project for launches.

Inside are stored:

• LAUNCH_STRATEGY.md - the logic of launching and key stages.

• KEY_MESSAGES.md - key meanings that need to be conveyed.

• OBJECTIONS.md - a list of typical objections and types of materials that respond to them.

Claude:

• does not contradict itself from post to post

• takes into account what stage of warming up you are currently at

• helps scale content without losing overall logic.

Example Task

Write a post for the middle of the warm-up: without direct selling, but with an increase in trust and value. Rely on the current status from LAUNCH_STRATEGY.md and meanings from KEY_MESSAGES.md.

As a result, the launch looks coherent. Less overheating of the audience, easier to engage helpers: everyone has one source of truth regarding the logic of the launch.

How to Build Your First Skill: Step-by-Step Route

Even a basic skill already provides a noticeable time saving. The process can be broken down into 4 steps.

Step 1. Find a Repeating Task

First, it would be good to ask yourself: what are you explaining to the AI over and over again?

In my case, it is:

• template for posts /videos /letters/ ad creatives

• analysis procedure (channels, competitors, reviews)

• format of internal documents (orders, memos, protocols)

• logic of warming up and launches.

Step 2. Describe the Process in Human Terms

At this stage, you need to document:

• the goal, what is considered a good result

• the steps, what comes after what

• mandatory elements (structure, blocks, constraints)

• examples of successful and unsuccessful outcomes

• typical mistakes.

At this stage, a single Markdown file with clear text without code is sufficient.

Step 3. Package into the Structure of the Skill

The minimum set of files looks like this:

my-skill/
  SKILL.md            # main skill file: brief description + when to use + basic rules
  instructions.md     # (optional) detailed instructions and examples (if you don't want to inflate SKILL.md)
  examples/           # (optional) examples of inputs/outputs
  scripts/            # (optional) scripts for complex operations

The skill (Skill) is a folder with files. The main file in it is SKILL.md: it briefly describes what the skill does and when to apply it. The other files (instructions.md, examples/, scripts/) are optional, they are needed for detailed instructions, examples, and automation.

Here are my examples:

SKILL.md

---
name: telegram-post-writer
description: |
  Writes posts for Telegram in the author's style: calm tone, short paragraphs, without hype and "geekiness".
  Apply when asked: "make a post", "format for Telegram", "rewrite in my style", "write an analysis/explanation".
---

# Telegram Post Writer

## Goal
To create a post in "my tone": neutral, clear, practical.

## Basic Rules
- 1–2 emojis maximum
- paragraphs of 1–3 lines
- no "overhype/imbalance/urgent for everyone"
- if there are facts/numbers — ask for a link to the source or mark "needs confirmation"
- at the end: 1–2 conclusions "what this gives" + 1 nuance/limitation (if any)

## Formats (choose based on the task)
### News
1) Context (1–2 lines)  
2) What has changed (in list format)  
3) What this gives in practice (1–2 examples)  
4) Nuance/limitation  

### Service Analysis
1) What it does in simple words  
2) Who it is useful for (2–3 types of people)  
3) How to try it (3–5 steps)  
4) Nuances/pitfalls  

## Detailed Instructions and Examples
See instructions.md and the examples folder

instructions.md :

# Detailed: style and quality check

## Checklist before the final
- Is there specificity (example/step/case), not just general words?
- Is there no advertising/calls to action like “buy/subscribe”?
- Is it not overloaded with terms? If a term is needed — explain it briefly.
- If numbers/limits/dates are mentioned — is there a reference or note “needs confirmation”.

## Stop-words
“imba”, “razzia”, “urgent for everyone”, “exploded the market”, “cannon”, “has no analogues”. 

examples/input.md:

Create a post for Telegram about the news:
“Service X added function Y. Now you can …”
Tone: calm analysis, without hype.
Length: up to 1200 characters.

examples/output.md:

Service X has introduced function Y — it has become easier to do Z without manual routine.

What has changed:
• …
• …

Where it will be useful in practice:
— …
— …

Nuance: …

Step 4. Test and refine

Run the skill on real tasks and see:

• whether it activates correctly (not too often and not too rarely)

• whether the result meets your expectations

• what specific cases need to be added.

As a result, clarify the instructions and examples, thus refining the Skill.

Step 5. At your discretion, you can share with the team

In the Team /Enterprise plans, skills can be shared across the organization. What does this provide? For example, new employees can immediately work according to the correct templates. The quality of documents and analysis becomes more uniform, employees do less “their own way” and more according to the standard.

What distinguishes Skills from other capabilities of Claude

To avoid confusion, it is useful to distinguish three things: Skills, projects, and MCP-connectors.

1) How Skills differ from Projects

• Projects are a set of files and context that is always loaded while you are in that project.

• Skills are procedures that are activated only when the task is truly about them.

Use projects for storing knowledge bases and artifacts, and skills for describing recurring processes.

2) How Skills differ from User Instructions

• User instructions operate globally: "always write this way", "address in 'you'" etc.

• Skills are specific to certain contexts: "when we create a post for the channel - here are the rules", "when we analyze channels for procurement - here are the criteria".

In this way, User instructions are global general preferences for everything. Meanwhile, skills are used for specific scenarios.

3) How Skills differ from MCP connectors

MCP (Model Context Protocol), whose source code is also open, gives Claude access to external tools and data:

• workspaces in Notion;

• tasks in Jira /Asana;

• chats in Slack, etc.

Skills answer the question "how do we use this": how exactly we format notes about calls in Notion, how we create tasks, how we write in-chat updates.

That is, MCP provides access, while Skills provide the process and standards.

Limitations and pitfalls

Skills have their own "buts" that are important to remember:

• Skills are not suitable for constantly changing processes. If you experiment with the format every day and have not yet determined what is normal and what is not, it will be difficult to maintain a skill. It’s better to first find a stable scenario.

• The quality of the skill depends on the quality of the instructions. Here, as they say, "you reap what you sow". If you provide vague formulations, contradictions, and do not give examples, the result can be unpredictable.

• Limitations based on tariffs and security policies. Skills are currently only available on paid plans. Plus, not every organization allows running third-party code, even if it resides within the skill.

Should you start using Skills

You should definitely consider using Skills if you regularly explain the same processes to AI, have stable formats: for posts, videos, documents, reports. You work in a team and want junior colleagues to rely on packaged experience. And of course, you are using a paid plan of Claude and are ready to spend time documenting processes.

You can wait if you are just getting acquainted with AI and still don't understand what tasks you want to delegate. Your processes are currently chaotic and change significantly from week to week. Plus, if the company has strict limitations on running code and storing instructions in external services, it’s better to manage without them.

The most effective approach is not to try to describe everything at once. Take one task that really annoys you with its repetitiveness, build a skill around it, and observe how this skill works for about two weeks. If you feel the effect, scale the approach to other scenarios. You can support me on my Telegram channel, where I write about what I understand or am trying to figure out, testing useful AI services, tools for office, business, marketing, and video.