Index live · v1.7.0 · NOV 18 2026

The Manifest · Issue 010 ·

A scanner said DO NOT INSTALL on shadcn/ui. We read all 158 findings.

How to vet a skill before you install it, three MCP servers we ran before listing them, and a one-star skill that made the cut.

Hi again.

This week: what a skill can run on your machine before the model reads a word of it, and the ten checks we’d do before installing one; what a security scanner told us about two skills in our catalog, and why we wouldn’t trust either label; three MCP servers we ran before listing them; and a skill with one GitHub star that earned its place.

Same promise as always: what passed, what didn’t, and what we’re learning about which skills actually work. No generic AI news roundup. Just useful stuff.

THE SIGNAL / 01

A skill can run shell commands before the model reads it, and a scanner score won’t tell you which ones.

Three features turn a skill from a text file into code on your machine. allowed-tools pre-approves tools for the turn that invokes it; Claude Code’s docs say it “does not restrict which tools are available.” A !`command` line runs before the skill is sent to Claude, and “never prompt[s] for permission.” Outside auto mode, a command that isn’t approved aborts the skill, but a skill can pre-approve its own command. And a skill’s hooks keep running for the rest of the session with your full user permissions.

So we ran NVIDIA’s SkillSpector 2.13.4 on two skills in our catalog, static analysis only. shadcn/ui scored 100 out of 100: DO NOT INSTALL, 158 findings. We read the flagged lines. 123 came from one rule about unpinned commands and 23 from the spacing in Markdown tables; the seven high-severity ones pointed at CLI docs, a config table, a comment and a React example. We saw no sign of malicious intent, though that isn’t a full audit. Differential Review scored 7: SAFE, one finding. Its report never mentioned that it pre-approves Bash with no scope, the first thing we’d check.

Two skills aren’t a benchmark, and the scanner’s optional AI review didn’t run. The point isn’t that SkillSpector is wrong. A score is a list of places to look, not a verdict.

What to do with this:

1. Before you install a skill, read its frontmatter. Treat allowed-tools as permission for those tools to run without asking you.

2. Search the SKILL.md for !` lines, and for code fences opened with !. Each one runs before the model reads the skill.

3. Read the instructions themselves, not only the scripts. Snyk found its 76 malicious payloads in skills’ markdown instructions.

NEW & VETTED / 02

Three MCP servers we installed and ran before listing them.

  • Busabaseis a local database and workspace your agents share. One command starts it, with no account or API key, and agents connect over MCP. We had one Claude Code session create a table and three records, then asked a fresh session what was in it, and it answered from Busabase. Writes arrive as change requests, but in a default local workspace ours merged straight away, so don’t read a merged change as a person’s approval. MIT, 759 stars.
  • OrkasVideoStudiogives your agent tools to render motion graphics, edit footage you supply and assemble both from an editable plan. We asked for a 15-second explainer with no keys, and got a playable video; its own checks caught several mistakes along the way. It isn’t on npm yet, so you install it from source, about 465 MB with dependencies. MIT, 346 stars.
  • Coding Tools MCPgives a chat app such as Claude Desktop file search, multi-file patches, shell commands and git inside one project folder. On our test bench, an agent used only the server’s own tools to write a small Python module with a failing test, fix it with a patch and rerun the test. Then it tried five ways to reach outside that folder, with absolute and ../ paths through both the file tools and the shell, and every one was refused. We didn’t try symlinks. Two things to know: commands run for real inside the folder, and it sends anonymous usage counts to PostHog unless you set DO_NOT_TRACK=1. Apache-2.0, 1,225 stars.

THE PICK / 03

Visualizer — Lex from SigmaStudioz

B-RANK · 1 STAR · 1 CONTRIBUTOR · UPDATED NOVEMBER 1, 2026

It had one GitHub star when we listed it, so stars told us nothing; we tested it instead. Visualizer turns a topic, a model or a dataset into something you can use or watch: an interactive diagram, a data explorer, a Three.js scene, or an explanation video with captions and editable sources.

What earned it the spot: it gives an agent a capability it doesn’t have on its own. Its bundled script renders a storyboard to video, and it tells the agent to keep real evidence apart from illustrative inputs and to report what it verified. Our explainer and video both rendered.

The caveats: video needs FFmpeg and Pillow, and an optional sound helper fetches licensed audio from Openverse. On one of our machines the soundtrack was mono and inaudible through a USB output; a stereo copy played fine. We don’t know why.

$ npx skills add sigmastudioz/agent-skills@visualizer

DIDN’T PASS / 04

One server we didn’t test.

  • Screenpipe’s MCP servernot listed. It searches what you’ve seen, said and heard, from a recorder that can capture your screen, audio and keyboard input. Its docs say that data stays local, but testing it means recording your machine, and we didn’t. Its source is free for personal use and needs a commercial license for work; the official app is sold separately.

WORKFLOW / 05

Security Code Review

Trail of Bits’ security skills on every pull request, with static analysis, a review focused on what changed, and GitHub connected. Intermediate, about 10 minutes to set up.

It includes Differential Review, the skill from 01 that pre-approves Bash. That suits a security reviewer that needs to run tools, but it means the turn you invoke it, shell commands your deny and ask rules don’t catch run without asking. That same turn reads the pull request, so text someone plants in a diff could steer one of those commands. Read its frontmatter first, then decide.

WHAT CHANGED / 06

53 skills you may have installed were updated since Issue 009.

“Updated” means the skill’s own directory got a commit since our check on November 2, not a judgment about what the commit did. Before you run update, read the commit.

Seventeen are Microsoft’s Azure and Foundry skills, changed together on November 9. Four are Superpowers (Brainstorming, Systematic Debugging, Subagent-Driven Development and Executing Plans), changed together on November 3. Four are Anthropic’s document skills (PDF, PPTX, DOCX and XLSX), changed together on November 6, and four are the .NET Foundation’s. Top of the rest by stars:

…and 15 more. Each skill’s page shows when it was last updated.

That’s it for Issue 010. See you next week.

— The Manifest

Back to the archive

Newsletter

Stay ahead of the AI skills curve

Every issue: what we added to the directory, what we rejected and why, and one workflow worth stealing. Nothing you'd get from an AI news roundup.

Free forever. Unsubscribe anytime. No spam. Read a past issue.