claude-onboarding-agent
Interactive onboarding agent that configures your Claude Code workspace for optimal use. Through guided questions, it generates tailored CLA
Open source Open in the app JSON README (API)
About
Interactive onboarding agent that configures your Claude Code workspace for optimal use. Through guided questions, it generates tailored CLAUDE.md and AGENTS.md files and installs the right skills for your workflow — whether you're coding, researching, writing, or managing projects.
Details
- Kind
- Plugins
- Topic
- No topic detected
- Publisher
- a2ngerer
- Origin
- marketplace
- Category
- ferramentas
- Stars
- 5
- Last push
- 2026-09-07T12:36:35Z
- Repository state
- ativo
- Language
- Shell
- License
- MIT
- Added
- 2026-08-30 01:48:58
- Updated
- 2026-08-30 01:48:58
- Origin id
a2ngerer/claude_onboarding_agent/claude-onboarding-agent
README
<p align="center">
<img src="docs/assets/banner.jpeg" alt="Claude Onboarding Agent" width="720">
</p>
<h1 align="center">Claude Onboarding Agent</h1>
<p align="center">
<strong>Claude is powerful. But only if set up right.</strong><br>
From blank project to a Claude Code setup tailored to <em>your</em> work — in under five minutes.
</p>
<p align="center">
<img src="docs/assets/demo.gif" alt="Claude Onboarding Agent demo" width="820">
</p>
---
Most people start a new Claude session and just... start typing. No context, no workflow, no structure. Results are inconsistent, Claude forgets everything between sessions, and it never really learns how you work.
This plugin fixes that. Run `/onboarding` once — Claude scans your project, asks a handful of targeted questions, and generates everything you need: a tailored `CLAUDE.md`, subagent roles, tool permissions, workflow instructions, and more.
Already know what you need? Call any setup skill directly.
---
## Pick your path
Not sure where to start? Run `/onboarding` and we'll figure it out with you. Or jump in:
| You are… | Run this | You get |
|---|---|---|
| **A developer** shipping code | `/coding-setup` | Superpowers workflow, subagent roles, stack permissions |
| **Building a web app** (frontend, backend, or full-stack) | `/web-development-setup` | Frontend, backend, or full-stack web app — framework-aware permissions, env-var hygiene, deploy-target pointers |
| **A data scientist / ML engineer** | `/data-science-setup` | Notebook hygiene, experiment tracking, `data/raw→processed` layout, reproducibility |
| **Building a personal wiki / second brain** | `/knowledge-base-setup` | OKF v0.1 (Karpathy-pattern) wiki, optional Obsidian CLI subagent |
| **Writing emails, memos, reports, proposals** | `/office-setup` | Business-writing focus: Q1 branches guidelines (email path vs. report path); presentations out of scope |
| **A researcher or academic** | `/research-setup` | Literature review workflow, Zotero reference management, reading notes |
| **Writing a thesis, paper, or dissertation** | `/academic-writing-setup` | Thesis / paper / dissertation — LaTeX or Typst, Zotero, citation rules that prevent hallucinations |
| **Creator brand voice** (YouTube, shortform, newsletter, podcast) | `/content-voice-setup` | Voice + audience + per-platform rule files — scope limited to writing guidance; publishing / analytics / media production out of scope |
| **Running infra / DevOps** | `/devops-setup` | Cloud + IaC + CI config, safe-by-default infra workflow |
| **Designing UIs** | `/design-setup` | Design tool + frontend stack, UI guidelines, no generic AI looks |
| **Already set up, but Claude burns tokens searching large repos / docs / PDFs** | `/graphify-setup` | Local [Graphify](https://github.com/safishamsi/graphify) knowledge-graph index + `/graphify` slash command + PreToolUse hook (consulted before Grep/Glob/Read) — layers on top of any other setup |
---
## Install
### Option 1 — Plugin marketplace (recommended)
This repo ships a self-hosted `marketplace.json`, so the standard plugin marketplace flow works directly:
```
/plugin marketplace add a2ngerer/claude_onboarding_agent
/plugin install claude-onboarding-agent@claude-onboarding-agent
```
### Option 2 — One-liner
**macOS / Linux**
```bash
curl -fsSL https://raw.githubusercontent.com/a2ngerer/claude_onboarding_agent/main/scripts/install.sh | bash
```
**Windows (PowerShell)**
```powershell
irm https://raw.githubusercontent.com/a2ngerer/claude_onboarding_agent/main/scripts/install.ps1 | iex
```
Clones the repo and links all skills into `~/.claude/skills/` so Claude Code picks them up automatically. To update: re-run the same command.
> How it works under the hood, plus what the future plugin path looks like: [docs/installation.md](docs/installation.md)
### Uninstall
**macOS / Linux**
```bash
curl -fsSL https://raw.githubusercontent.com/a2ngerer/claude_onboarding_agent/main/scripts/uninstall.sh | bash
```
**Windows (PowerShell)**
```powershell
irm https://raw.githubusercontent.com/a2ngerer/claude_onboarding_agent/main/scripts/uninstall.ps1 | iex
```
---
## What's inside
| Command | What it does |
|---|---|
| `/onboarding` | Orchestrator — scans your repo, infers your use case, routes you to the right setup |
| `/coding-setup` | Installs [Superpowers](https://github.com/obra/superpowers), wires up brainstorm → plan → subagents → review → commit |
| `/web-development-setup` | Framework-aware web-app setup — Next.js / React / Vue / Svelte / SolidJS / Astro / Remix + optional backend (Node/Bun/Python/Go). API conventions, component structure, env-var hygiene, deploy-target pointers |
| `/data-science-setup` | Notebook workflow (Jupyter/marimo), experiment tracking (MLflow/W&B/DVC), reproducible `pyproject.toml`, `data/raw/interim/processed` layout, model-card pointers |
| `/knowledge-base-setup` | Builds a [Karpathy-pattern](https://github.com/forrestchang/andrej-karpathy-skills) wiki from your notes or codebase, conformant to Google's [Open Knowledge Format (OKF v0.1)](https://github.com/GoogleCloudPlatform/knowledge-catalog/blob/main/okf/SPEC.md) so the bundle stays portable across OKF-aware agents (+ optional [Obsidian](https://obsidian.md) CLI integration via dispatched subagent — no always-on MCP token cost) |
| `/office-setup` | Business writing — emails, memos, reports, proposals. Q1 bifurcates emitted guidelines (email path / report path / both); presentations are out of scope |
| `/research-setup` | Input side of academic research — literature reviews, paper screening and summaries, Zotero reference management, reading notes. Manuscript drafting lives in `/academic-writing-setup` |
| `/academic-writing-setup` | Thesis / paper / dissertation setup — LaTeX or Typst stack, Zotero + Better BibTeX, citation style, no-invented-citations rules, `sections/`/`bib/`/`figures/` scaffold |
| `/content-voice-setup` | Creator brand voice for YouTube / shortform / newsletter / podcast. Emits per-platform rule files under `.claude/rules/` (`youtube.md`, `shortform.md`, `newsletter.md`, `podcast.md`) keyed to the platforms Q1 selects. Scope limited to writing guidance — publishing, analytics, thumbnails, and media production are out of scope |
| `/devops-setup` | Cloud provider, IaC tool, CI/CD — safe infra workflow + agent roles |
| `/design-setup` | Design tool, frontend stack, accessibility standard — UI guidelines without the generic AI look |
| `/graphify-setup` | Installs [Graphify](https://github.com/safishamsi/graphify) (25-language tree-sitter + Markdown + PDF + media indexer). Registers `/graphify query / path / explain` and a PreToolUse hook consulted before file-search tool calls — cuts token cost on large codebases and mixed-media corpora. Safe to layer on top of any other setup |
| `/checkup` | The single maintenance entrypoint. Runs the audit, weighs findings + meta age + deprecated-model anchors, and either prints a short "fine-as-is" summary, hands off to `/upgrade-setup`, or invokes `/onboarding --rebuild`. Call this whenever you want to know whether your setup still matches current best practices |
| `/anchors` | Refresh anchor-derived marker sections in CLAUDE.md/AGENTS.md against the latest upstream anchors |
| `/audit-setup` | Read-only power-user tool — normal flow is `/checkup`. Audits your existing Claude setup (permissions, CLAUDE.md quality, git hygiene, tooling) and returns a prioritized HIGH/MEDIUM/LOW findings list without changing any files |
| `/upgrade-setup` | Power-user / internal tool — normal flow is `/checkup`. Re-applies current best practices to an existing setup: per-change diff preview, dry-run flag, timestamped backups, never touches content outside the plugin's delimited sections |
---
## How it works
```
/onboarding
│
▼
Scan repo — detect files, manifests, existing CLAUDE.md
│
▼
Suggest the most likely use case (or ask if the repo is empty)
│
├── 1. Coding Setup
├── 2. Web Development
├── 3. Data Science / ML
├── 4. Knowledge Base Setup
├── 5. Office & Business
├── 6. Research & Writing
├── 7. Academic Writing
├── 8. Content Creation
├── 9. DevOps & Infrastructure
└── 10. Design & Frontend
│
▼
Ask 3–7 targeted questions
│
▼
Install Superpowers (always for Coding/KB, optional for others)
│
▼
Generate CLAUDE.md + config files
│
▼
Print a completion summary
```
---
## What gets generated
Every setup skill creates a tailored `CLAUDE.md` with context and instructions specific to your workflow. Here's what each path produces:
| Skill | CLAUDE.md | Agents | settings.json | .gitignore | Hooks | External |
|-------|-----------|--------|---------------|------------|-------|----------|
| Coding | ✓ + workflow | ✓ 3 roles (AGENTS.md) + opt-in `code-reviewer` subagent | ✓ stack permissions | ✓ stack | — | Superpowers + opt-in GitHub MCP |
| Web Development | ✓ + pointers (`.claude/rules/api-conventions.md`, `component-structure.md`, `env-vars.md`) | opt-in `component-auditor` subagent | ✓ framework + package-manager + deploy-CLI permissions | ✓ `node_modules/`, framework build outputs (`.next/`, `dist/`, `.astro/`, …), `.env.local`, test artifacts | ✓ type-check on save (TS only, opt-in) | Superpowers (optional) + opt-in GitHub MCP |
| Data Science | ✓ + pointers (`.claude/rules/data-schema.md`, `evaluation-protocol.md`) | opt-in `notebook-auditor` subagent | ✓ uv / notebook / tracker permissions | ✓ raw data, notebook checkpoints, experiment artifacts | ✓ nbstripout on save (opt-in) | Superpowers (optional) |
| Knowledge Base | ✓ + Karpathy / OKF v0.1 bundle (`wiki/` + `index.md` / `log.md`) | ✓ `.claude/agents/obsidian-vault-keeper.md` (optional) | — | ✓ | — | Superpowers + Karpathy |
| Office | ✓ + writing style | — | — | ✓ | — | Superpowers (optional) + opt-in Gmail / Calendar / Drive MCP |
| Research | ✓ + citation format | — | — | ✓ LaTeX | — | Superpowers (optional) |
| Academic Writing | ✓ + non-negotiable citation rules (pointers to `.claude/rules/writing-style.md`, `citation-rules.md`) | opt-in `writing-style-auditor` subagent | — | ✓ LaTeX / Typst build artifacts | ✓ SessionStart rules reload (opt-in) | Superpowers (optional) |
| Content | ✓ + brand voice | — | — | ✓ media files | — | Superpowers (optional) |
| DevOps | ✓ + infra workflow | ✓ 3 roles | ✓ stack permissions | ✓ IaC state, secrets | ✓ plan-before-apply guard (opt-in) | Superpowers (optional) |
| Design | ✓ + UI guidelines | ✓ 2 roles | ✓ stack permissions | ✓ design assets | — | Superpowers (optional) + opt-in Figma MCP |
| Graphify | ✓ + delimited pointer block (`/graphify query / path / explain`) | — | Graphify registers its own PreToolUse hook via `graphify install` | — | (Graphify's own PreToolUse hook) | `graphifyy` (Python package via `uv tool install`, `pipx` fallback — note: the PyPI package name is `graphifyy` with a double y; the CLI command and the [GitHub repo](https://github.com/safishamsi/graphify) use `graphify` with a single y) |
### The coding workflow (powered by Superpowers)
The Coding Setup installs [Superpowers](https://github.com/obra/superpowers) — a widely used, battle-tested Claude Code workflow library — and wires it into your `CLAUDE.md`. Every future session follows a proven loop:
```
Brainstorm idea → Write plan → Dispatch subagents → Code review → Commit
```
### The knowledge base (Karpathy pattern)
The Knowledge Base Setup sets up the [Karpathy LLM Wiki pattern](https://github.com/forrestchang/andrej-karpathy-skills): a `raw/` folder for source material and a `wiki/` folder of interlinked markdown notes that Claude builds and maintains. Drop files into `raw/`, ask Claude to ingest them — the wiki grows automatically.
The generated `wiki/` is an [**OKF v0.1 bundle**](https://github.com/GoogleCloudPlatform/knowledge-catalog/blob/main/okf/SPEC.md) — Google Cloud's Open Knowledge Format, the published, vendor-neutral standard that formalizes the LLM-wiki pattern. Each note is one concept with YAML frontmatter (a required `type` plus recommended `title` / `description` / `tags` / `timestamp`), reserved `index.md` / `log.md` files, and standard Markdown cross-links (not wikilinks). Conformance keeps the knowledge base portable across any OKF-aware agent or tool, not just Claude.
Optional: connect [Obsidian](https://obsidian.md) via the official Obsidian CLI, wired into a dedicated `obsidian-vault-keeper` subagent. Obsidian renders OKF frontmatter as native Properties and Markdown links in its graph view, so the bundle stays both OKF-conformant and fully navigable. Vault reads/writes only load the CLI reference when actually invoked, so chats that don't touch the vault pay zero Obsidian tokens — unlike a persistent MCP whose tool schemas are loaded into every session.
---
## Realtime anchors
Some best practices change faster than plugin releases — current Claude model IDs, recommended MCP servers, Claude Code tooling conventions, subagent patterns, knowledge-base layouts. To keep skills accurate without requiring users to reinstall the plugin, this repo ships a set of short, auto-updated reference snapshots under [`docs/anchors/`](docs/anchors/).
Skills fetch anchors at runtime from a pinned `raw.githubusercontent.com` URL via the shared [`skills/_shared/fetch-anchor.md`](skills/_shared/fetch-anchor.md) protocol (24h cache, embedded offline fallback). A daily GitHub Action researches the sources listed in each anchor and opens a PR for human review — anchors are never auto-merged. Format details: [`docs/anchors/README.md`](docs/anchors/README.md).
---
## Skill trigger evals
Skill routing depends entirely on each skill's frontmatter description. To catch silent routing regressions, every skill ships a trigger-eval fixture under [`evals/`](evals/) — realistic should-trigger and should-not-trigger prompts. The runner asks a judge model to route each prompt against the live catalog and reports per-skill pass rates:
```
scripts/run-skill-evals.sh # full suite (claude CLI or ANTHROPIC_API_KEY)
scripts/run-skill-evals.sh --skill coding-setup # one suite, e.g. after a description change
```
CI runs the full suite weekly (`skill-evals.yml`); fixture presence and validity are enforced on every PR. Design: [`docs/superpowers/specs/2026-06-12-skill-eval-harness-design.md`](docs/superpowers/specs/2026-06-12-skill-eval-harness-design.md).
---
## Language support
All skills detect your language automatically from your first message and respond accordingly. Supported: English, German, Spanish, French, and any other language Claude Code supports.
---
## Contributing
See [CONTRIBUTING.md](CONTRIBUTING.md) for how to add new setup skills.
---
## Authors
- Alexander Angerer — <alexander.angerer@outlook.de>
- Maximilian Achenbach — <Maximiliana28@gmail.com>
---
## License
[MIT](LICENSE) — free to use, modify, and distribute.