Skills are packaged playbooks (instructions, templates and scripts) that Claude loads only when a task calls for them. This guide covers what they are, why they're the highest-leverage feature in Claude right now, how to turn them on for any plan, and how to write one that actually fires.
--- name: weekly-status-update description: Build the Tuesday client status email (RAG, top 3, risks, capacity). Use when asked for a status update, sprint recap or "what do we tell the client." --- # Weekly Status Update ## Inputs - Client name + sprint number - Asana board (from the client charter) ## Steps 1. Pull completed + in-flight tasks 2. Flag anything past due as a risk 3. Fill templates/status.md 4. Keep it under 300 words
Every team using AI hits the same wall. The output is only as good as the prompt, and the prompt is only as good as whoever is typing it that day. So the senior person rewrites the brief, re-pastes the brand guide, re-explains the format, and quietly becomes the bottleneck all over again.
Skills move that knowledge out of people and into the system. Write the playbook once, and Claude applies it every time the task shows up, for anyone on the team.
Without skills: the loop above, every single time.
With skills: the process runs itself.
A skill is a directory with a SKILL.md file at its core, plus any templates, reference docs or scripts it needs. The trick is progressive disclosure: Claude doesn't load everything up front. It reads only as deep as the task requires, so you can install dozens of skills without bloating every conversation.
Anthropic skills. Built in for Excel, Word, PowerPoint and PDF. Claude uses them automatically when you ask for one of those files.
Custom skills. Yours. Your process, your templates, your standards, written in Markdown.
Org skills. Team and Enterprise owners publish or provision skills to everyone, optionally on by default.
Partner skills. Built by tools like Notion, Figma and Atlassian to pair with their connectors.
The format, tone, checks and guardrails are written down, so output stops depending on who typed the request or how tired they were.
No pasting brand guides or re-explaining templates. Ask for the deliverable and the playbook is already loaded.
Institutional know-how becomes an asset the whole team uses, and it survives turnover, vacations and growth.
Only a one-line description sits in context until a skill is needed, so you can stack many without crowding out the work.
The same skill runs in the Claude apps, Cowork, Claude Code and the Office add-ins. Skills also follow an open standard (agentskills.io).
Teams can share, review, publish, require and security-scan skills, so the good process spreads and the bad one doesn't.
The most common mistake is using the wrong tool for the job: stuffing a process into custom instructions, or expecting a connector to know how your team works. Here's the split.
| Feature | What it gives Claude | When it loads | Use it for |
|---|---|---|---|
| Skills | Procedures: how to do a specific task, with templates and scripts | Only when the task matches the description | Repeatable deliverables and workflows |
| Projects | Static background knowledge and files | Always, inside that project | A body of context for one initiative or client |
| Custom instructions | Broad preferences about you and your style | Every conversation | Tone, role, formatting defaults |
| Connectors (MCP) | Access to external tools and data | When Claude needs that system | Reading or writing in HubSpot, Asana, Slack, Drive |
| Plugins | A bundle of skills, connectors and commands | Installed as a package | Shipping a whole role's toolkit at once |
The combo that wins: connectors give Claude access, skills give it judgment. A HubSpot connector lets Claude read your deals. A pipeline-review skill tells it which deals matter, what "at risk" means at your company, and how to write it up.
Pick the problem you're trying to solve.
We'll tell you which Claude feature solves it and why.
Skills need code execution and file creation turned on, since they run in Claude's code environment. The steps depend on your plan.
Official reference: Use Skills in Claude (Anthropic Help Center).
One skill, one job. "Weekly client status email" is a skill. "Help with clients" is not.
State what it does and when to use it, with the actual phrases people will type. This is what Claude matches against.
Inputs, steps, output format, guardrails. Plain Markdown. Add an example of good output.
Templates, a REFERENCE.md, or scripts. Reference each file from SKILL.md so Claude knows when to open it.
The folder itself must be the ZIP root, and the folder name should match the skill name.
Enable it, run several prompts that should trigger it, and check Claude's reasoning to confirm it loaded.
The description is the single most important line in your skill. Draft yours here and watch the checks.
↑ Copy this, fill in the brackets, save it as SKILL.md in a folder with the same name.
Not everything should be a skill. Check what's true about the task you have in mind.
Score your task against the signals that make a skill pay off.
We run our delivery on skills. These are real examples from our library, plus what each one replaced.
Colors, fonts, tone, logo rules and layout patterns, applied to any deck, doc or page Claude builds.
Replaces: "make it look like ours" rework on every deliverable
One source of truth per engagement: stakeholders, Asana board, Slack channels, capacity, guardrails. Every other skill reads it.
Replaces: re-explaining the client in every chat
RAG status, hard-date table, top 3, risks and capacity check, assembled from live project data in a fixed format.
Replaces: an hour of copy-paste per client, per week
Reviews the call transcript, checks every action item against the board, and proposes the missing tasks and follow-ups.
Replaces: action items lost between the call and the board
Turns a sprint from the roadmap spreadsheet into properly formatted tasks with the right fields, owners and scope.
Replaces: vague tickets devs can't act on
A weighted score from Slack, Asana and call data that flags churn risk before the client says anything.
Replaces: gut-feel account reviews
The pattern: the most valuable skills aren't clever prompts. They're your operating procedures written down, chained together, and pointed at your real systems. That's where AI goes from a chat toy to an operating layer.
Most "Claude ignored my skill" problems are really description problems. Most "the output was off" problems are missing examples.
Skills load automatically when your request matches. You can also name the skill in your prompt to force it.
Install organization skills from the directory (Customize, then "+", then the Skills tab) and let Claude chain them across your connected tools.
Skills enabled in your Claude account sync into Claude Code when you sign in with the same account. Run /skills to see them. Sync is one-way and doesn't apply to API-key sign-in.
In the Claude add-ins, type / to pick a skill or just describe the task. Templates in a skill can load straight into the open file.
Available in beta through the code execution tool. Dependencies must be pre-installed in the container, since packages can't be installed at runtime.
Bundle skills with connectors and commands into a plugin to roll out a complete role-based toolkit in one install.
Share from the three-dot menu with specific people (or groups on Enterprise). Shared skills are view-only, stay grayed out until the recipient turns them on, and update automatically at next use. Access ends when someone leaves the org.
"Publish to org" puts a skill in the shared library as available to install, installed by default, or required. With review on, it moves through pending, changes requested and published, and updates go back through review before anyone gets them.
Enterprise orgs can scan third-party skills before they run. Scans usually take a minute or two. A blocked skill has to be fixed and re-uploaded.
Uploaded skills are private until shared. Admins see names and sharing status but not file contents, and owners can turn off user-created skills or gate them with custom roles.
Skills can include scripts and install packages in Claude's environment. The two real risks are prompt injection (instructions that steer Claude somewhere you didn't intend) and data exfiltration (code or instructions that send your data out).
Work through the list before you toggle a third-party skill on. This applies to skills from colleagues too.
Code execution isn't on. Individuals: Settings > Capabilities. Team and Enterprise: an owner has to enable it in Organization settings > Plugins & skills.
Check four things in order: the skill is toggled on; the description says when to use it with phrases that match how you asked; the instructions are clear; and if all else fails, name the skill in your prompt. Then go back and tighten the description so you don't have to.
The usual culprits: the ZIP is over the size limit, the folder name doesn't match the skill name, there's no SKILL.md file, the files sit loose at the ZIP root instead of inside the folder, or the name or description has invalid characters.
Code execution is disabled, either for your whole organization or for your account. Skills shared with you also appear grayed out until you turn them on yourself.
Your org owner may have turned off user-created skills, or your custom role is missing the "Create skills" capability.
The sharing toggles may be off, your role may lack the capability, or the skill isn't one you created. Built-in skills, skills shared with you and skills installed from the library can't be published.
Skills are folders of instructions, scripts and resources that Claude loads on demand to do specialized tasks well, such as building branded documents, following a reporting format, or running a team-specific workflow.
Free, Pro, Max, Team and Enterprise. Code execution and file creation have to be on.
No. A skill can be a single Markdown file with a name, a description and instructions. Scripts are optional and only worth adding when a step has to run the same way every time.
It reads each enabled skill's name and description and loads the full instructions only when your request matches. That's why the description, which should say what the skill does and when to use it, matters more than anything else in the file.
A project holds background knowledge that's always loaded inside that project. A skill holds a procedure that loads only when the task calls for it, and it works across all your chats.
Yes, paired with a connector. The connector gives Claude access to the system, and the skill tells it how your team uses that system: which fields matter, what the naming conventions are, and what "done" looks like.
Skills you write yourself, or get from trusted sources and review, are low risk. Third-party skills can contain scripts or instructions that cause prompt injection or data exfiltration, so read every file before you enable one, and use security scanning if you're on Enterprise.
JayOh designs, writes and deploys Claude skills wired into your CRM, project tools and reporting, so your team's best process runs every time without you in the loop.
Talk to JayOh
We engineer growth systems. JayOh builds the RevOps architecture, GTM execution and data infrastructure that turn messy revenue operations into predictable pipeline for B2B SaaS and PE-backed teams.
