# How To Add Skills to Claude AI - Text Humanizer > [!question] Lingering Questions > - Does Claude AI vet an uploaded skill's instructions before it runs, or is the upload flow trust-on-install with no review step? > - The transcript garbles the demo repo's name ("Blader Humanizer") via auto-caption — verify the actual GitHub repo before recommending this specific skill to anyone. > [!tip] Article Ideas > - None — this is a short screen-recording how-to, not analytical source material. > [!link] Cross Reference > - [[3. Resources/Permanent Notes/A Shared Kernel Is Not a Trust Boundary for Untrusted Agent Code]] — adjacent theme: installing unvetted third-party instructions into an agent's context. > - [[AI Agent Safety MOC]] — the no-vetting-shown install flow is a small, concrete instance of the supply-chain-trust question this MOC tracks. > [!quote] Vocabulary > - **Skill (Claude AI product sense)** — a packaged set of instructions and optional files that Claude loads to perform a specialized task on demand, installable via drag-and-drop `.md` or `.zip` upload with no coding required. ## Bib capsule Visualmodo (YouTube), "How To Add Skills to Claude AI: Text Humanizer," https://www.youtube.com/watch?v=GsLk-9QrS54 ## Author's thesis (Adler Q1+Q2) A short screen-recording tutorial demonstrates the mechanical path for installing a third-party Agent Skill into Claude AI's web app: find a skill on GitHub, download its `.zip`, and upload it through Claude's Personalize → Connectors and Skills panel. In detail: the panel supports three ways to add a skill (write your own, have Claude help write one, or upload an existing package), and once installed a skill exposes a toggle plus a replace/edit/try/uninstall menu — but the video shows no step where the installer reviews the skill's actual instructions for safety before enabling it. ## Key arguments 1. Adding a Skill to Claude AI is a pure download → upload → toggle flow requiring no code. 2. The demo installs a third-party skill sourced from a public GitHub search with no vetting step shown — trust is implicit in the install action itself. ## Claim atoms ### Claude AI installs skills by uploading an .md or .zip > [!quote] ID:rw_pending > "You can upload in .md or .zip, okay? Click on drag and drop files, select the skill that you just have downloaded, click on open." **Context & use**: [00:01:09]–[00:01:42] — Personalize/Customize → Connectors and Skills → Skills → plus → upload. Useful mechanical reference for how a packaged skill gets distributed and installed. **Tags**: #xref **Highlight**: [view highlight](readwise://pending) <!-- atomicity: 1 idea ✓ --> ### A skill toggle and three-dots menu manage the skill lifecycle > [!quote] ID:rw_pending > "click into these three little dots and uninstall this. You can replace, you can edit, you can try it out or uninstall." **Context & use**: [00:01:42]–[00:02:13] — enable/disable toggle plus replace/edit/try/uninstall, all reversible from the same panel. **Tags**: #xref **Highlight**: [view highlight](readwise://pending) <!-- atomicity: 1 idea ✓ --> ## Permanent candidates - None. This is a short procedural demo with no reusable, cross-project claim that survives the atomicity + reusability filter — the deep-read notes reach the same conclusion. The "no vetting step shown" observation is real but too thin on its own (single anecdotal video, no counter-case or mechanism) to stand as an atomic claim; it's logged as a Lingering Question instead. ## Processing checklist - [x] Verbatim quotes carried over from the existing deep-read notes - [x] Checked for a Permanent-note-worthy claim — none found; resonance filter applied per source's own thinness - [x] Connect into the relevant MOC — [[3. Resources/MOCs/AI Agent Safety MOC]] ## Links - [[3. Resources/MOCs/AI Agent Safety MOC]] — the unvetted third-party skill install is a small instance of this MOC's trust-boundary tracking - [[3. Resources/Readwise/Articles/How to Add Skills to Claude AI Text Humanizer]] — source document - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — applies-in: Joins the AI tooling for knowledge work cluster.