Setup Fassi Skills

---
name: setup-fassi-skills
description: Install my curated agent skills into the current project, with recommendations based on what the project is.
disable-model-invocation: true
---

Install skills from the **manifest**, the `manifest.json` beside this file, into the current project. The manifest is the single source of truth: every skill name, install command, category, workflow and tool comes from it.

## 1. Read the manifest and the project

Read the manifest. Then fetch the published one, https://raw.githubusercontent.com/sh4wty1/fassi-skills/main/skills/setup-fassi-skills/manifest.json, and compare the skill names. When the published manifest has names this one lacks, list them and give the command that updates this copy: `/plugin marketplace update fassi-skills` when this skill runs as `/fassi-skills:setup-fassi-skills`, `npx skills@latest update -g` otherwise. Carry on with the manifest beside this file either way. When the fetch fails, carry on without comment.

Survey the project from its root: `README*`, `AGENTS.md` / `CLAUDE.md`, the package manifest and the top-level layout.

A manifest skill is **installed** when it sits in a `*/skills/` directory of the project, or when this session already lists it (from a plugin or a global install).

The project root is the nearest ancestor holding `.git` or `package.json`; one installer writes there, so run every command from it. When neither marker exists, say so and ask whether to continue in the current directory.

Done when you hold two sets, neither holding an installed skill, each skill with a one-line reason drawn from what you read:

- **recommended**: the skills whose `when_to_use` fits this project, one workflow chosen per `workflow_rule`.
- **worth knowing**: at most 5 skills that fit only in part, where the project could use the skill but nothing you read calls for it. The reason says what would make it fit.

An empty or unreadable project yields two empty sets.

## 2. Ask how to proceed

Open with the installed skills in one line, when there are any. Show the recommended set and then the worth-knowing set, each with its reasons. When the recommended set is empty because the skills that fit are already installed, say that rather than "nothing to recommend". Then ask one question with these options:

- **Install recommended**: the recommended set alone, without the worth-knowing set. Offer it first, and only when the set is non-empty.
- **Install all**: every skill in the manifest. Put the count in the option, since each installed skill adds its description to every session.
- **Choose myself**: go to the picker.
- **Help me decide**: ask in plain text what they are building, plus at most two follow-ups (new or existing project, feature size, whether PRs are reviewed on GitHub). Rebuild the recommended set from the answers, show it, and go to the picker.

Ask with your structured question tool when you have one (`AskUserQuestion` in Claude Code); otherwise print a numbered list and read the numbers back.

## 3. Picker

First ask which categories to browse: one multi-select whose options are the manifest categories, each with its skill count, suffixed `(Recommended)` when it holds a recommended skill. Then one multi-select question per chosen category. Each option is a skill: its name as the label, its `when_to_use` as the description. Suffix the label with `(Recommended)` for the recommended set, `(Worth knowing)` for the worth-knowing set and `(installed)` for installed skills.

`AskUserQuestion` takes 1-4 questions per call and 2-4 options per question, so:

- the category list, and a category with more than 4 skills, split into `<category> (1/2)`, `(2/2)`;
- a category with 1 skill merges into its nearest neighbour;
- more than 4 questions continue in a second call.

### Suggestions from outside the manifest

When the project needs something no manifest skill covers, look for a **suggestion**: first among the other skills of each source's `repo`, then with `npx --yes skills@latest find <keyword>`. Keep at most 4, each one read (its `SKILL.md`) before you offer it.

Suggestions get their own picker question, `suggestions`, with the repo URL in each description so the user can review it. They are installed only when picked there: "Install recommended" and "Install all" cover the manifest alone.

## 4. Install

The target **agents** default to the agent running this skill. Mention that default in one line with the question in step 2, so the user can name others.

For each chosen skill, in manifest order, one at a time: take its source's `install` command (for a suggestion from an unlisted repo, `any_repo_install` with `{repo}` as `owner/name`), replace `{skill}` with the skill name and `{agents}` with the space-separated agent ids, and run it from the project root. A skill marked `(installed)` is reinstalled only when the user picked it.

Agent ids are the installer's own, and the installers disagree on a few (`gemini-cli` / `gemini`, `kilo` / `kilocode`, `kiro-cli` / `kiro`; the `impeccable` source wants `claude` / `cursor`, not `claude-code`, and takes them comma-separated in `--providers`). When an installer rejects an id, read the supported list in that source's `repo` and retry with its spelling.

A skill counts as installed when a `skills/<name>/SKILL.md` exists under an agent directory of the project afterwards (`.claude/skills/`, `.agents/skills/`, `.cursor/skills/`, ...). Judge by that file, since an installer can exit 0 having installed nothing. Keep going after a failure.

## 5. Report

A table with every chosen skill: `installed` with the path, `already present`, or `failed` with the last lines of the installer's output. Then, in a few lines:

- the workflow these skills form and the order to use them;
- each entry of `tools` with its `run` command;
- for each suggestion installed, the ready-to-paste `skills` entry (plus a `sources` entry when its repo is new) to add to `manifest.json` in https://github.com/sh4wty1/fassi-skills, so it becomes part of the registry;
- any need still uncovered.