Back to search
skillselect
MarketplaceWorks with Claude CodeGitHub
Reads every skill you have installed, scores each one against your request, and tells you which to use, when and how. A Claude Code plugin.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Checked today
Badge for your README
Markdown
[](/listings/alibb007--skillselect)HTML
<a href="/listings/alibb007--skillselect"><img src="/api/listings/alibb007--skillselect/badge.svg" alt="findsafeskills safety rating" /></a>Install
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 18 lines
- {
- "name": "skillselect",
- "owner": {
- "name": "Ali B.B",
- "url": "https://github.com/alibb007"
- },
- "plugins": [
- {
- "name": "skillselect",
- "source": "./",
- "description": "Tells you which of your skills to use, when and how, finds more on GitHub, and never installs anything.",
- "version": "3.0.0"
- }
- ],
- "metadata": {
- "description": "SkillSelect: which skill to use, when and how, for Claude Code."
- }
- }
README.md · README · 154 lines
- <p align="center">
- <picture>
- <source media="(prefers-color-scheme: dark)" srcset="assets/hero-dark.svg">
- <img src="assets/hero-light.svg" alt="SkillSelect: which of your skills to use, when and how. Finds more on GitHub. Never installs anything." width="100%">
- </picture>
- </p>
- **You install many skills and Claude uses one.** SkillSelect is a Claude Code plugin that reads every skill on your computer, scores each one against what you asked for, and tells you which to use, when and how. It also finds skills on GitHub that fit your request. It never installs anything.
- <p align="center">
- <img src="assets/demo.gif" alt="Film: installed skills pile up. A person asks Claude to build a landing page and Claude picks one skill. SkillSelect reads all 50 installed skills, scores them, shows the 4 that fit with when and how to use each, and finds 5 more on GitHub. Then token usage, time and steps, with and without SkillSelect." width="100%">
- </p>
- <p align="center"><sub>The film is drawn from saved runs in a test folder with 50 public skills. Asked to build a landing page, plain Claude loaded 1 of the skills that fit in 5 of 6 runs. SkillSelect listed 4 installed skills and 5 from GitHub, each with when and how to use it. <a href="assets/demo.mp4">Download the clip</a>.</sub></p>
- <p align="center">
- <a href="LICENSE"><img alt="License: MIT" src="https://img.shields.io/badge/license-MIT-2A78D6?style=flat-square"></a>
- <img alt="Claude Code plugin" src="https://img.shields.io/badge/Claude_Code-plugin-101A2B?style=flat-square">
- <img alt="Python 3, standard library only" src="https://img.shields.io/badge/python-3_·_standard_library_only-178A63?style=flat-square">
- <img alt="No telemetry" src="https://img.shields.io/badge/telemetry-none-5B6A80?style=flat-square">
- </p>
- > - **Problem.** People install dozens of agent skills. Asked for a job, the assistant loads one and leaves the rest unused.
- > - **What I built.** A two-stage retrieval pipeline in plain Python: query expansion, BM25 with field weights, rank fusion of fit and GitHub stars, then a re-rank step by the model.
- > - **How I tested it.** I wrote the pass rule before the run, tuned on 12 requests, froze the code with a checksum, and ran 24 sealed requests once.
- > - **Result.** On the sealed requests, GitHub suggestions that do the job went from 40% to 88%, and installed suggestions from 60% to 100%. One real request used 91% fewer tokens than asking without the plugin.
- ## Start in 3 steps
- **1. Install it with one line.** This needs Claude Code 2.1.275 or later.
- ```
- /plugin install skillselect --marketplace alibb007/skillselect
- ```
- On an older version, use two lines:
- ```
- /plugin marketplace add alibb007/skillselect
- /plugin install skillselect@skillselect
- ```
- **2. Type what you want to build.**
- ```
- /skillselect build a landing page
- ```
- **3. Read which of your skills fit, and when and how to use each.** You also get up to 5 more from GitHub, each with the command to add it. You run that command yourself.
- A second command sorts the skills you already have into categories:
- ```
- /skillselect categorize
- ```
- It needs Python 3. SkillSelect never uploads your skills and never installs or deletes anything. There is no login, no paid API and no telemetry. It contacts GitHub only: the search words go to api.github.com, and public skill files are read from raw.githubusercontent.com. If the `gh` command is signed in or `GITHUB_TOKEN` is set, the finder sends that token to GitHub's API with its searches, as `gh` itself would.
- To try it without installing, clone the repository and start Claude Code with the folder:
- ```bash
- git clone https://github.com/alibb007/skillselect
- claude --plugin-dir skillselect
- ```
- <picture>
- <source media="(prefers-color-scheme: dark)" srcset="assets/flow-dark.svg">
- <img src="assets/flow-light.svg" alt="When you have a job: you type /skillselect and the job, SkillSelect scores your skills, adds GitHub skills, and keeps what does the job with when and how for each. On every request: one sentence is added for you, and you get a short answer you can use as it is." width="100%">
- </picture>
- ## How it picks
- <picture>
- <source media="(prefers-color-scheme: dark)" srcset="assets/pipeline-dark.svg">
- <img src="assets/pipeline-light.svg" alt="Retrieval pipeline in four lanes. LLM layer: request, query expansion. Local index: term model, candidate gate, BM25F score, top 8. Remote index: GitHub search, fetch skills, gate and BM25F, rank fusion, top 12. LLM layer: re-rank, answer. Counts on the arrows: 50 skills, 5 pass, 73 repos, 106 files, 30 pass." width="100%">
- </picture>
- Read it as a U. The request goes in at the top left and the answer comes out at the bottom left. The counts on the arrows are from one run of "landing page" in the 50-skill test folder. 5 of the 50 installed skills passed the gate. On GitHub, 73 repositories led to 106 skill files, and 30 of them passed.
- 1. **Query expansion.** Claude turns your request into 2 to 4 topic words and up to 6 close words: other names for the same thing, and its file types.
- 2. **Term model.** Words are cut to their stems. Two words side by side also count as a pair, so "landing page" is a term of its own. Topic words and pairs weigh 1, and close words weigh 0.5.
- 3. **Candidate gate.** A skill becomes a candidate when it shares two topic words with the request, or one topic word and one close word, or three close words. The gate skips skills. It never skips your request. A skill that shares one common word, such as "page", would only add noise.
- 4. **Fit score.** Each candidate gets a BM25 score. A rare word counts more than a common one, repeating a word helps less each time, and a word in the skill's name counts 3 times.
- 5. **GitHub.** The finder sends up to 13 searches to the GitHub API. It opens the repositories that mention your topic and the largest skill collections, and reads the skill files inside them. Those skills pass the same gate and get the same score.
- 6. **Rank fusion.** Each GitHub skill has a place by fit and a place by stars, and the two places are fused into one number. A famous repository with a weak fit cannot lead, and a perfect fit that nobody has starred cannot lead either. At most 2 skills per repository are listed.
- 7. **Re-rank.** Claude reads the top 8 installed and the top 12 GitHub descriptions. It keeps at most 5 from each list, and only skills whose own description says they do the job.
- ```
- fit(skill) = Σ weight(term) × idf(term) × tf × 2.5 / (tf + 1.5 × (0.25 + 0.75 × len / avg))
- idf(term) = ln(1 + (N − df + 0.5) / (df + 0.5))
- order(skill) = 1 / (60 + place by fit) + 1 / (60 + place by stars)
- ```
- BM25 with field weights follows Robertson and Zaragoza (2009). Reciprocal rank fusion with k = 60 follows Cormack, Clarke and Büttcher (2009). The design is the usual retrieve, then re-rank. The finder is Python 3 with the standard library only, and the code holds no word list or tuned number.
- ## Tested
- ### Does it list skills that do the job?
- <picture>
- <source media="(prefers-color-scheme: dark)" srcset="assets/lift-dark.svg">
- <img src="assets/lift-light.svg" alt="GitHub picks that do the job on 24 sealed requests: plugin 2.2.0 40% (26 of 65), with the retrieval math 40% (40 of 101), with the LLM re-rank 88% (49 of 56). Installed picks that do the job: 12 of 20, 14 of 28, 14 of 14. Installed fits found: 12, 14 and 14 of 19." width="100%">
- </picture>
- Claude Opus wrote 36 everyday requests without seeing any skill name or description. 12 were for tuning and 24 were sealed. I set the pass rule first: at least 10 points more GitHub picks that do the job, with no loss on installed skills. Before the sealed run I froze the code and wrote its sha256 into the rules file. The sealed 24 ran once.
- An AI judge, Claude Opus, labelled which skills do each job. It never saw which version had listed a skill.
- The new version passed. Job by job it listed more GitHub skills that do the job in 13 of the 24 requests, the same number in 9 and fewer in 2. The rules, the requests, the lists, the labels and the scripts are in [evidence/round9](evidence/round9).
- ### What one request costs
- <picture>
- <source media="(prefers-color-scheme: dark)" srcset="assets/numbers-dark.svg">
- <img src="assets/numbers-light.svg" alt="Token usage: 528,087 without SkillSelect, 49,854 with it, 91% fewer. Time to get the answer: 1 min 51 s without, 51 s with, 54% less. Steps to find your skills: 27 without, 1 with, 96% fewer." width="100%">
- </picture>
- I ran `/skillselect build a landing page for my product` three times in the test folder, each on an empty cache. The runs took 32, 51 and 86 seconds and about 49,900 tokens each, and all three listed the same 9 skills. The picture shows the middle run.
- The side without the plugin is one run from 18 Nov 2026. Plain Claude was asked which skills to use and how, with web search and the `gh` command. It took 27 steps, 111 seconds and 528,087 tokens. Both sides are in [evidence/round10](evidence/round10).
- I also asked plain Claude to build the landing page in that folder six times, without the plugin. It loaded 1 skill in five runs and 2 in one ([evidence/round7](evidence/round7)).
- ### Short answers
- SkillSelect also adds one sentence to every request, so the answer comes back short and finished. In a test of 42 requests the answers had half the words (245 against 488), and none had a blank left to fill in (0 of 42 against 6 of 42).
- <picture>
- <source media="(prefers-color-scheme: dark)" srcset="assets/proof-dark.svg">
- <img src="assets/proof-light.svg" alt="Words per answer: 488 for a plain request, 245 with SkillSelect. Answers with a blank: 6 of 42 plain, 0 of 42 with SkillSelect. Output tokens per answer: 1,946 plain, 1,447 with SkillSelect." width="100%">
- </picture>
- ## Limits
- - The math alone did not raise the share of GitHub picks that do the job on the sealed requests. It was 40% before and 40% after. It found more good candidates (40 against 26), and the re-rank step turned them into the 88%.
- - The judge and the re-rank step are the same model, Claude Opus, so the 88% and the 100% may flatter it.
- - 5 of the 19 installed skills that fit were still missed.
- - Claude Opus wrote the 36 test requests, and I chose the fields they cover.
- - The side without the plugin is one run, and the three runs with it all used one request.
- - The seconds depend on GitHub. In the three runs the finder call alone took between 9 and 62 seconds.
- - The finder runs when you call it by name. Asked in plain words, Claude does not always start it.
- - "Needs" lists what a skill's instructions name, found with fixed patterns. It can miss a requirement or list an optional one.
- - Nothing is installed for you. Read a skill before you add it, because it is text your assistant will trust.
- - Without a GitHub token, GitHub allows 60 API calls an hour, which is enough for a few requests.
- ## License and credit
- SkillSelect is free to use, copy and change under the [MIT license](LICENSE). Keep the license notice with its copyright line when you copy or reuse it. A mention of SkillSelect or of Ali B.B with a link to this repository is welcome too.
- The evidence files quote short descriptions of public skills so the tests can be checked. Each one is named with its repository, and those texts belong to their authors. This repository does not include any of those skills.
- Claude and Claude Code are trademarks of Anthropic, and GitHub is a trademark of GitHub, Inc. SkillSelect is an independent project. It is not affiliated with or endorsed by either company.
- The software comes without warranty, as the license says. The numbers on this page are from the tests in [evidence](evidence), and they describe those tests only.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.