01 / 05
Relevance相关
How much of the skill's vocabulary actually shows up in your prompts. Rare-but-on-topic terms (like rebase) count more than common filler (like file).
relevance = Σ idf(t) · 1[t ∈ your prompts] / Σ idf(t)
Paste a Claude Code skill link. SkillsVote scores it against your local usage history and your actual machine — then tells you whether it's worth installing, and writes a one-click install prompt adapted to your setup. Nothing leaves your laptop.
uvx --from "https://github.com/horton2048/skillsvote-local/releases/download/v0.2.1/skillsvote-0.2.1-py3-none-any.whl" skillsvote
Matches 38 prompts in your history · keywords: commit, rebase, branch, PR · last used 2 days ago · you don't have this skill yet.
A SKILL.md file teaches a coding agent (Claude Code, Codex, OpenClaw …) how to do one specific job — write a migration, scaffold a Next.js page, debug a flaky test. There are now over 1.68M of them on GitHub. Browsing them feels like browsing the App Store with the reviews removed.
A skill is a markdown file with front-matter, a description, and instructions. The agent reads it before tackling a task. Think of it as a tiny, focused expert that the agent loads on demand.
--- name: git-cli-power-user description: Run git workflows from the CLI: branching, rebasing, conflict resolution, PRs. when_to_use: user is in a git repo and asks about commits, branches, or pull requests. --- # Instructions 1. Always check `git status` before suggesting fixes. 2. Prefer `gh pr create` over web flows…
For every skill you actually use, you'll evaluate ten. The "is this for me?" decision is annoying because the answer depends on three things only you know:
SkillsVote runs locally as a tiny HTTP server on port 8773. The whole pipeline below happens between hitting Enter and seeing your score.
A GitHub URL (folder or file), a skills.vote page, or a raw SKILL.md. SkillsVote fetches it and reads the front-matter.
Reads your ~/.claude prompt history, tokenizes it, builds an IDF index, and probes your shell for the binaries the skill needs.
Five dimensions roll up into a single 0–100 value, with a verdict and a copy-paste install prompt already adapted to your machine.
Scroll. Each panel shows what one dimension actually measures — and what it sees on a real machine.
How much of the skill's vocabulary actually shows up in your prompts. Rare-but-on-topic terms (like rebase) count more than common filler (like file).
How often you actually do tasks in this skill's domain. A log-saturated count of your prompts that hit the skill's distinctive terms — so a million mentions of "the" don't inflate the number.
When did you last do something this skill could've helped with? Exponential decay with a half-life of 14 days, so a hot 3-day-old hit weighs more than a stale six-month-old one.
High demand × not-already-owned = a real capability gap. If you've already installed this exact skill (or a near-duplicate slug), Gap is heavily discounted — installing twice doesn't help.
Will it actually run? Checks your OS against any platform markers in the skill, and probes for required binaries on your PATH. A perfectly relevant skill that needs Docker on a Docker-less laptop drops here.
A weighted blend of all five. SkillsVote turns that into one of four verdicts — install, optional, skip, or already installed — plus a one-paragraph "why".
SkillsVote is one Python process bound to 127.0.0.1. There is no telemetry, no analytics SDK, no "anonymous" payload — and the network only gets touched at one well-defined moment.
The HTTP server only listens on loopback. Your LAN can't reach it, let alone the internet.
~/.claude stays putPrompt history is tokenised and IDF-indexed in memory. The corpus is read at startup; never copied, never written elsewhere.
The only network request is when you paste a link — SkillsVote fetches that skill's public README. Nothing else is uploaded.
~2,000 lines of Python. Read every line on GitHub before you run it.
Every assessment ends with one of four labels — so you can triage a long list of candidate skills the way you triage a PR queue.
Relevant, in demand, recent activity, and your environment supports it. Worth the context-window slot.
You touch this area sometimes. Install if you're stocking up the library; skip if you're keeping it lean.
Either you don't do this kind of work, or your machine can't run it. Bookmark and re-check later.
Matches a slug already in your installed skill set. SkillsVote flags it and suggests an upgrade-or-skip.
No. SkillsVote-local is the scoring tool — it runs entirely on your machine without any API key. The optional hosted skills-vote integration (a separate package from MemTensor) is the only thing that uses a key.
From ~/.claude on your machine — Claude Code stores your prompt sessions there. You can point at a non-default path with --claude-home PATH. If you've never run Claude Code, there's simply no history to score against and SkillsVote will say so.
The scoring model is agent-agnostic — it just needs a prompt corpus. The current scanner reads Claude Code's format; PRs adding readers for Codex, OpenClaw, etc. are very welcome on GitHub.
Trust your gut over the number. SkillsVote is a triage signal, not a verdict you have to obey — it tells you why it scored what it did, and you decide. The output shows the matched terms, prompt count, and recency so you can sanity-check.
skills.vote (and the upstream MemTensor/skills-vote) is a global recommendation index — what's popular, what's high-quality across all users. SkillsVote-local is the personal layer on top: given a candidate, is it for me, on this machine?
Yes — paste a path to a local SKILL.md or a private GitHub URL (with GH_TOKEN in your environment). The fetch happens on your machine; nothing about the skill is shipped anywhere.
One command. No account. No telemetry. Closes when you close the terminal.
uvx --from "https://github.com/horton2048/skillsvote-local/releases/download/v0.2.1/skillsvote-0.2.1-py3-none-any.whl" skillsvote
pip install it.