There's nothing to memorize. Skills aren't commands. When the moment calls for it, Claude pulls them out on its own. Just talk the way you normally would.
Skip this and Claude won't know it's installed. The skill list is read once, at startup. Close the window you installed in and open a new one.
If a list comes back, you're done. If nothing shows up, you either haven't restarted, or the install didn't finish → see the Install guide again.
They're not a button you press, Claude reaches for one on its own when it decides it's needed. So just talk the way you normally would. You don't need to memorize any names.
You may not notice right away. It won't reach for one on a short question — only when the task grows does it pull one out.
As of the current release, there are 48 skills. If you bought a single group, only that group is installed.
Same work, fewer tokens. Makes the biggest difference when your usage limit feels like it's draining fast.
Try saying
“Keep your answers short” · “Find the over-engineered parts of this repo” · “Compress CLAUDE.md”
ponytail |
Builds only as much as needed, kept short. Filters out unnecessary work first |
ponytail-audit |
Scans the whole repo and ranks the most over-engineered spots |
ponytail-review |
In changed code, only what to delete is targeted — reinvented standard features, unused dependencies, speculative abstractions |
caveman |
Compresses answers aggressively. 65% fewer output tokens (measured by the author). Technical content stays intact |
caveman-commit |
Shortens commit messages. Keeps the meaning and the reason |
caveman-compress |
Compresses CLAUDE.md and note files to cut the input tokens. Code, paths, and structure are preserved |
caveman-review |
Code review comments, one line each — location, problem, fix |
investigate-first |
Narrows down the cause before fixing. For failures with no clear cause, or flaky errors |
lean-build |
When building a new feature, how far to build before stopping is decided first |
safe-refactor |
Keeps behavior unchanged and only cleans up structure. Verifies before and after |
surgical-patch |
Fixes only at the narrowest layer — so nothing next to it breaks |
verify-and-stop |
Proves it's done, then stops. Doesn't touch extra things along the way |
Makes Claude check its own work while coding, narrow down causes, and plan before it starts.
Try saying
“Find the root cause of this bug first” · “Review this code before I merge it” · “Let's write the tests first”
requesting-code-review |
Before merging, makes it check itself: did it actually do what was asked |
systematic-debugging |
When a bug shows up, narrows the cause before fixing it |
test-driven-development |
Locks in behavior with a test first, then writes the code. When a bug is reported, reproduces it first |
karpathy-guidelines |
Cuts down on mistakes LLMs commonly make while coding — over-engineering, over-broad fixes, hidden assumptions |
code-review-and-quality |
Runs code review from several angles before merging |
code-simplification |
Keeps behavior the same, only trims for readability |
debugging-and-error-recovery |
Digs down to the root cause when a build or test breaks |
performance-optimization |
Looks at frontend, backend, queries, and database performance together |
security-and-hardening |
Hardens user input, login, storage, and external integrations |
archify |
Turns architecture, flow, and sequence diagrams into one clickable HTML page |
security-audit |
Full-repo security review. Built by Cloudflare |
tirith |
Checks terminal commands before they run — supply-chain attacks, hidden code |
agents-best-practices |
Designs, audits, and explains AI agent architecture |
Changes how it works. It plans without being asked, and actually runs things before saying it's done.
Try saying
“Let's figure out what to build first” · “Actually run it to confirm it's done” · “Write the plan down first”
brainstorming |
Before building, what to build gets clarified first. Always runs before creative work |
verification-before-completion |
Before saying "it's done," it makes sure the work actually runs . |
writing-plans |
Writes the plan down before touching code |
incremental-implementation |
When touching multiple files, feeds changes in small pieces instead of dumping them all at once |
planning-and-task-breakdown |
Breaks a big job into ordered tasks. For when you don't know where to start |
spec-driven-development |
When requirements are fuzzy, writes a spec before starting |
skill-inspector |
Checks whether a new skill is safe before installing it. Built by NVIDIA |
loopy |
Turns a repeated task into a loop, saves it, and reuses it |
icm-architect |
Designs work and knowledge as a folder structure so Claude doesn't lose track on long tasks |
Strips the "made by AI" look. Picks color, font, and spacing with a reason behind each choice.
Try saying
“Strip the AI-made look off this page” · “Clean up the color and font on this screen”
ui-ux-pro-max |
Gives you the whole decision base for building screens — 67 styles, 161 palettes, 57 font pairings, 25 charts |
design-taste-frontend |
Keeps landing pages and portfolios from looking like a template . |
high-end-visual-design |
Locks in the rules behind why expensive-looking sites look expensive — font, spacing, shadow, card structure |
hallmark |
Catches and fixes the AI-made look. Can also pull a design system from a URL or screenshot |
baoyu-design |
Turns mockups, prototypes, landing pages, and dashboards into a single HTML file. |
Offloads work that isn't code, too.
Try saying
“Write something up from this” · “Make this document easier to read”
humanizer |
Smooths out AI-sounding sentences. Meaning stays the same |
video-shotcraft |
Builds product videos shot by shot from a script |
codex-ppt |
Turns writing, reports, and papers into a visually consistent PPTX deck |
avoid-ai-writing |
Finds and fixes "AI phrasing" in writing |
Repetitive work like documents and spreadsheets.
Try saying
“Clean up this table” · “Summarize this file into a document”
obsidian-bases |
Handles notes like a spreadsheet with Obsidian Bases (.base) |
obsidian-cli |
Reads, writes, and searches an Obsidian vault from the command line |
officecli |
Creates, edits, and proofreads Word, Excel, and PowerPoint files |
Researches a topic and hands back the results with sources.
Try saying
“Research this topic and put it together”
book-to-skill |
Turns a book or PDF into a skill — pulling out its framework and principles |
anysearch |
Live web search and pulling the body text from a URL |
ponytail-audit skill to scan this repo”, like that.
The name is in the left column of the table above.--check to the end of the command you were given and run it again.
Without installing anything, it just shows what's currently installed .~/.claude/skills — just delete that folder inside it.
Everything else keeps working.If installation fails or a skill isn't recognized, just tell us.