Short description:

Install once. Six AI coding tools share a common set of rules and safety checks for every project.

Product introduction:

The Minimal Vibe Coding Kit is a set of files. You add it to your project once. This file set informs the AI coding tools how your project runs and how to work safely.

The AI coding tools are code-writing programs that assist you. Claude Code and Cursor are two examples. The kit supports six tools: Claude Code, Cursor, Codex, OpenCode, Grok, and Kimi Code. The kit works with all programming languages.

Why you should use it:

• AI coding tools write code very quickly, but they can guess wrong. They may modify too many files. They might forget your rules in a new working session.
• The kit provides the tools with a project map and clear boundaries. You receive better results and fewer unintended changes.

How it works (about 2 minutes):

1. Install the kit with one command (see the "Installation" section below).
2. Open your project in your AI coding tool. Paste a prompt.
3. Review the changes the tool suggests. Respond "yes" to agree.

Then the tool writes the backbone.yml file. This file acts as the project map. The tool reads it at the start of each working session.

What’s included in the kit:

• Rules: short limits that the tool always reads.
• Skills: pre-written workflows. The tool only loads a skill when needed for a task. The kit has 26 skills.
• Commands: short words like /security-scan to quickly run common skills.
• backbone.yml: a common project map for all six tools.

Skills to try first:

Clarifying requirements

• prompt-sharpener: transforms a rough idea into an exact requirement and implements it immediately.
• clearthought and sequential-thinking: breaks a large or vague task into smaller steps you can verify.

Working with checks

• clean-delivery: makes a small change through six steps: specify, code, clean, architect, harden, and verify. Each step must pass a check before moving to the next.
• proofline-orchestration: uses AIs with different roles. One AI writes code. Another AI finds errors in the results. A test decides if it’s done or not.
• parallel-analysis: runs 2 to 5 read-only analyses (no file modifications) simultaneously, then combines and verifies the results. Suitable for large changes and questions about the entire project.
• autoresearch-coding: tests each small change one at a time. Only retains changes that produce a number you choose to be better.

Keeping the project safe

• agentshield-security-review: checks AI configuration files (rules, skills, hooks, and MCP) before merging. It looks for risks such as prompt injection.
• threat-model-security-review: reviews application code (API, login, data input methods) for security risks. It reports what it finds and what hasn’t been proven.
• path-sensitive-shell-safety: checks shell scripts using file paths. It ensures that an empty or overly broad path never reaches the rm command.

Daily support

• coding-level: choose the level of explanation from 0 (very simple) to 5 (expert). The kit retains your choice for future sessions.
• memento: saves notes at the end of a working session for the next session.
• wait-what: used when an answer isn’t clear. The tool rephrases it in simple language, according to your language.
• mermaid: draws flowcharts and system diagrams.

The kit also has 12 other skills, such as clone-website, visual-design-loop, and daily-workflow-curator. See the full list on GitHub.

Safety features:

• The kit does not overwrite your CLAUDE.md or AGENTS.md files. It only adds a marked block.
• All changes during installation await your approval.
• You can list protected paths in