A collection of AI agent skills, tools, and benchmarks for vibe coding workflows.
vibe-coding-tools/
├── agent-skills/ # Reusable agent skills
│ ├── claude-code-statusbar/ # Color-coded Claude Code status line + popup alerts
│ ├── distill-user/ # Mine transcript history into durable user-preference memory
│ ├── internet-search/ # Domain-first, route-independent web search (no API keys)
│ ├── markdown-render-qa/ # Catch Markdown that reads fine but renders wrong
│ ├── nano-banana-pro-vertex/# Cinematic hero-image generation (Gemini 3 Pro Image on Vertex AI)
│ ├── remove-background/ # Transparent PNG generation pipeline
│ ├── svg-diagram-qa/ # Render-and-inspect QA for hand-authored SVG diagrams
│ ├── vibe-deck/ # Automated PPTX & slide generation
│ └── anthropic-skills/ # Anthropics submodule library
└── ai-benchmark/ # AI code & reasoning benchmarks
These tools are designed for and tested with frontier-class AI models:
| Model | Provider | Notes |
|---|---|---|
| Claude Fable 5 | Anthropic | Highest-capability long-horizon agentic work |
| Claude Opus 4.8 | Anthropic | Deep reasoning + complex tasks |
| Claude Sonnet 5 | Anthropic | Strong planning + execution, daily driver |
| GPT-5 | OpenAI | Next-gen reasoning / Plan mode |
| Gemini 3.x Pro | Google DeepMind | Advanced agentic coding / Plan mode |
Note: These skills require models with tool-use capabilities (file viewing, code execution, image inspection). Smaller or non-agentic models will not produce equivalent results.
Reusable AI agent skills — drop into .agents/skills/ for any agentic coding assistant.
| Skill | Description |
|---|---|
| claude-code-statusbar | Color-coded Claude Code statusLine showing model, context size, usage bar, with escalating macOS popup alerts at 33%, 85%, 90%, and 95% context usage. |
| distill-user | Mine your local Claude Code transcript history for what you've said more than once — repeated requests, standing checks, corrections, preferences, dislikes — and distill it into memory rules with verbatim evidence, behind one approval gate. Runs entirely locally; workspace-scoped by default, global scan is opt-in. |
| internet-search | Domain-first, route-independent web search — prefers specific catalogs (GitHub, Stack Overflow, arXiv, Hugging Face, package registries) over generic web search, validates every hit with a live fetch, and fans out parallel workers for broad questions. Works with no WebSearch tool and no API keys (e.g. on a Vertex/Bedrock-blocked route). |
| markdown-render-qa | Catch Markdown that reads fine as raw text but renders wrong — collapsed setext headings, ineffective single-space hard breaks, **Field:** blocks merging into one paragraph — plus a per-case table for choosing Markdown vs HTML. Stdlib-only linter, no dependencies. |
| nano-banana-pro-vertex | Scriptable cinematic hero-image generation via Gemini 3 Pro Image (Nano Banana Pro) on Vertex AI — ADC auth, cost-tiered workflow (1K/2K/4K), parallel batch, PPTX-editable text-overlay sidecars. |
| remove-background | Remove white/light backgrounds from images to create transparent PNGs. Handles logos, icons, product photos, and multi-object images. |
| svg-diagram-qa | Render-to-PNG-and-inspect QA for hand-authored SVG diagrams — catches overlapping boxes, painted-over labels, off-canvas text, and viewBox cropping that well-formed XML hides. |
| vibe-deck | A Streamlined Solution for Generating Professional Slide PNG Images (Python/Pillow) with Optional PPTX Assembly, Speaker Notes, and Configurable Branding |
Included via git submodule from anthropics/skills. See their repo for license terms.
| Skill | Description |
|---|---|
| pptx | Create, edit, and read PowerPoint files via pptxgenjs or XML editing |
| docx | Create and edit Word documents |
| xlsx | Create and edit Excel spreadsheets |
| Parse, extract, and work with PDF files | |
| canvas-design | Design visual assets using HTML Canvas |
| frontend-design | Build polished frontend UI components |
| web-artifacts-builder | Create interactive web artifacts |
| webapp-testing | Test web applications systematically |
| algorithmic-art | Generate algorithmic and generative art |
| brand-guidelines | Create and apply brand identity systems |
| theme-factory | Build design themes and style systems |
| internal-comms | Draft internal communications and memos |
| doc-coauthoring | Collaborative document writing |
| mcp-builder | Build Model Context Protocol servers |
| skill-creator | Meta-skill for creating new skills |
| slack-gif-creator | Create animated GIFs for Slack |
Performance and cost comparison of Claude models on Vertex AI (Tests conducted October 2025).
Key Findings (Oct 2025):
- Haiku 4.5: ~2.3x faster, ~3x cheaper than Sonnet 4.5
- Sonnet 4.5: Superior reasoning for complex planning
Recommended Usage:
- 🎯 Plan Mode → Claude Opus 4.x (Best design, architecture, and reasoning)
- ⚡ Act Mode → Claude Sonnet 4.x (Ultra-fast implementation & solid reasoning)
Tooling:
- Google Antigravity
- Anthropic Claude Code
- Cline
See ai-benchmark/README.md for detailed results.
Provided "as is", used entirely at your own risk. These tools run shell commands, make network calls, and drive AI agents on your machine; some suggest permission-allowlist entries or wrapper scripts that reduce safety prompts.
- No warranty (as-is). All content in this repository — skills, scripts, wrappers, endpoint templates, configuration examples, and benchmarks — is provided "AS IS", without warranty of any kind, express or implied, including but not limited to merchantability, fitness for a particular purpose, and non-infringement (consistent with the MIT License).
- Limitation of liability. In no event shall the author or contributors be liable for any claim, damages, or other liability — including system failures, security breaches, data loss, or unintended actions taken by an AI agent or shell command — arising from or in connection with this software or its use.
- User responsibility. You bear the sole responsibility for evaluating safety, testing, securing your system and credentials, and complying with your organization's policies before use. Adopt only what you understand and accept, and prefer a real security boundary (least-privilege credentials, a sandbox/VM) for anything you cannot afford to have go wrong.
Tools for optimizing AI-assisted development workflows