UNKNWNAcceleration Create

From blocks 1 and 2. The lessons for blocks 3 and 4 are here.

Building skills in Claude, and the tips behind them

Thanks for building with me. This page has the slides, your assignment and twenty tips from the Claude docs and from my own work.

Title slide: Building skills in ClaudeOpen the slides20 slides, English

Your assignment: a skill from someone you learn from

Pick a person whose way of thinking you want on your side: a YouTuber, a blogger or an author. Collect their content, such as YouTube transcripts or the text of their blog posts. Ask the skill creator to make a skill from it. Test it on your own venture in a new chat, and bring one answer to show. Then ask the skill creator to evaluate your skill: does the answer get better with it than without it? Pick one person and one topic, not everything they ever said.

A skill is a folder with a SKILL.md

A skill is a folder with one file in it: SKILL.md. The top holds a name and a description. Below that go your instructions in plain markdown. The Claude docs give a simple test. If you keep explaining the same thing to Claude, write a skill. Good first skills are your commit message style, your brand voice, a review checklist or the way you write a weekly update.

The description does the work

Claude reads the description of every skill and matches your request by meaning. Say what the skill does and when to use it. Put the main use case first and use the words you type. "Helps with emails" is too vague. "Drafts replies in my voice. Use when I ask to write, answer or draft a mail" works. The name has lowercase letters, numbers and hyphens, up to 64 characters, and matches the folder name. The description can be up to 1,024 characters. A when_to_use field holds extra trigger phrases.

A skill is not CLAUDE.md, a hook or a subagent

CLAUDE.md loads in every conversation, so put rules there that always apply, like "never touch the database schema". A skill loads when your request matches it, so it holds the procedure for one kind of task. A hook runs when something happens, like a file save. A subagent works in a separate context and returns a result. MCP connects Claude to outside tools. Open your CLAUDE.md and move everything that only matters for certain tasks into a skill.

A skill is know-how, an agent puts it to work

My one-liners: a skill is a file the AI pulls from the cabinet. An agent is a colleague who gets to work with it. MCP is the power strip that connects the AI to the outside world. An orchestrator is the team leader who runs several skills, like the sprint-session skill from last week. A skill is passive: Claude loads it when needed. An agent is active: it decides, acts and tries again. Rule of thumb: know-how you want to reuse is a skill, a job that has to run on its own is an agent. They combine well, because an agent loads skills. The docs add one detail. A subagent does not see your skills unless you list them in its skills field.

Let one skill run the others

This is the orchestrator pattern. Give each skill one job and keep it small. Then write one more skill that is the team leader: its instructions say which skills to run, in which order, and what to hand from one to the next. My sprint-session skill works like this. It runs the vision board skill, then the user stories skill with that output, then the messaging canvas skill. My strategy-session skill lets five expert skills review each other's work until there is a homepage with copy. Start with two small skills that you already trust, and add the team leader last.

Keep SKILL.md short and put the rest in other files

Keep SKILL.md under 500 lines. Put detail in supporting files and link to them. Say when Claude should read each one. Use references/ for documentation, scripts/ for code and assets/ for templates and data. Claude reads a supporting file only when the task needs it. It runs a script without reading it, so only the output costs space. Point to the skill folder with ${CLAUDE_SKILL_DIR} and your paths always work.

Know where skills live and which one wins

A personal skill lives in ~/.claude/skills/name/SKILL.md and works in every project. A project skill lives in .claude/skills/ in a repo, and everyone who clones the repo gets it. A plugin can bundle skills. When two skills share a name, enterprise beats personal and personal beats project. Plugin skills have their own namespace, so they never clash. Pick specific names: frontend-review, not review. In the Claude app, open Customize, then Skills. That needs "Code execution and file creation" switched on in Settings, Capabilities.

Make your first skill in five minutes

In Claude Code, run mkdir -p ~/.claude/skills/weekly-update. Create SKILL.md in that folder and write a name, a description and your instructions. In the Claude app, start a chat and say "create a skill for my weekly team update". Describe the format and tone, and paste one real example you liked. Claude interviews you and builds the file. Click Save skill. The example in the Claude Academy has sections for wins, blockers and priorities, with at most three bullets each, in a direct and factual tone.

Let the skill creator write it with you

The skill creator is a skill that makes skills. It asks what the skill should do, when it should trigger and what good output looks like. Then it writes a draft, runs test prompts, shows you the results and rewrites. A script tunes the description so the skill triggers more often. Upload your best example of the output so it learns what good looks like. It also checks skills you already have. Ask it to evaluate a skill, and it runs the same test prompts with and without the skill, side by side, so you see whether the skill makes the answer better. Do that with a skill you installed from someone else too. In Claude Code you can install it with /plugin install skill-creator@claude-plugins-official.

Make a skill from someone you learn from

This is how I started. I loaded the books and articles of people I admire into Claude and asked it to work in their way of thinking. Pick one person and one topic. Collect their content: a YouTube transcript, or the text of their blog posts or newsletter. Give it to the skill creator and ask for a skill. Then ask the skill a real question about your own project and compare the answer with what you would have done without it.

Decide who can run it

Type /skill-name to run a skill yourself, with arguments after it if you like. By default you and Claude can both run a skill. Set disable-model-invocation: true and only you can run it. Use that for anything with side effects, like a deploy or a commit. Set user-invocable: false and only Claude can run it. Use that for background knowledge that should stay out of your slash menu.

Pass arguments

Use $ARGUMENTS for everything you type after the skill name. Use $0 and $1 for single words. You can also name them in the frontmatter, as in arguments: [component, from, to], and write $component in the text. Add an argument-hint like [issue-number] so autocomplete shows what to type.

Inject live data with a command

A line like !git diff HEAD in your skill runs before Claude reads the skill. The output replaces the line, so the skill gets fresh data every time. If a command can fail on purpose, add || true. One failing command stops the whole skill.

Run long jobs in a fork

Add context: fork and an agent (such as Explore, Plan or general-purpose) and the skill runs in its own subagent. Your main chat stays clean. The subagent sees only the skill text, not your conversation, so use this for skills with a clear task. A skill that only holds guidelines gives nothing back from a fork.

Limit what a skill may do

allowed-tools approves tools in advance while the skill runs, so Claude does not ask you each time. It does not block other tools. Approving Bash as a whole is broad, so narrow it, like Bash(git status *). disallowed-tools removes tools while the skill is active. paths limits automatic use to files that match a pattern. model and effort set the model and the effort for that skill.

Keep skills cheap

Claude always sees the description of every skill, so each skill costs a little. Keep descriptions sharp and bodies short, because every line of a loaded skill is a recurring cost. Write guidance as standing instructions, because Claude does not read the file again on later turns. Run /skill-doctor to find skills that cost space and never get used.

Test it like a normal user

Open a new chat. Ask for the task in your own words and leave out the skill name. Look for "Reading" and the skill name in Claude's thinking. Compare the answer with an example you like and sharpen every line you still fix by hand. To see what the skill adds, run the same task with and without it.

If it does not work, check this first

Does not trigger: improve the description and add the phrases you type. Does not load: SKILL.md must sit in its own folder, with exactly that name and valid YAML. Wrong skill fires: make the descriptions more distinct. Shadowed: the same name at a higher level wins, so rename yours. Runtime errors: install the packages the skill needs, run chmod +x on scripts and use forward slashes in paths. Run claude --debug to see errors and ask Claude which skills it has. Edits to an existing skill apply during the session. A brand new skills folder needs /reload-skills.

Find skills and check them before you trust them

skills.sh is a directory of skills, made by Vercel. You install one with npx skills add owner/repo, and the leaderboard ranks skills by installs. My rule of thumb: above 10,000 installs is a good sign, and a skill from Anthropic is safer than one from a stranger. Whatever the install count, open the repository, read the SKILL.md and look in the scripts folder. A Markdown viewer makes that quick: MarkEdit is free on the Mac, Typora is paid and also runs on Windows. The Claude docs agree. Install skills only from sources you trust, and review them first.

Share what works

Commit .claude/skills/ to your repo and your team gets the skills with the next pull. Put skills in a plugin to share them more widely. Skills follow an open standard, so outside Claude Code only a few fields work: name, description, license, compatibility, metadata and allowed-tools. Cloud sessions do not read your home folder, so enable the skill on your claude.ai account or commit it. Validate a skill before you hand it to anyone.