Prompt files are dependencies. Why aren't we versioning them?
If you use Claude Code, Cursor, or any agent that reads Markdown instructions, you probably have a folder of "skills": a deploy checklist, a code-review playbook, a set of rules someone shared on X. Look at where they came from. One is a gist, one was copied out of a GitHub repo three months ago, and one was pasted from Discord. You don't know which version you have, whether the author has fixed…
Prompt files are akin to dependencies in the realm of coding. Why isn't versioning of these files a standard practice? Claude Code, Cursor, and various agents that read Markdown instructions often rely on prompt files like skills - a deploy checklist, a code-review playbook, and rules shared on platforms like X. These files can come from various sources such as GitHub repos, Gists, or even Discord channels.
It's unclear which version you have, whether the author has made improvements since, or how changes would affect the agent if the file were re-pulled. This issue has been addressed for code decades ago with package managers, yet prompt files continue to be managed through copy-pasting.
To address this, Amitte was developed - an open registry for agent skills, agents, and MCP servers. Here's how it functions:
1. Each skill is versioned using the format publisher/name@version. Once published, the version remains static, with each updated version receiving a new unique hash (SHA-256). Modifying a version results in the publication of a new version, not a change to the existing one.
2. Bundles are used to pin the entire set of skills, not individual files. This feature emerged from a user who noted that versioning each skill separately was insufficient. In multi-skill projects, skills are designed to work together, and mismatched combinations have occurred. Bundle releases provide an immutable snapshot of the entire set, including the git commit it was created from. Users can install a bundle release by specifying the exact version, allowing them to roll back to a known-good state if needed.
3. The identity of the publisher is established through verification. Anyone can claim a domain name, but verified publishers receive a badge on their published items, serving a similar purpose to verified domains on npm or GitHub organizations.
4. The registry allows for search functionality. Semantic search enables users to find skills based on their purpose, even if the exact words do not match. An MCP server allows agents to perform their own searches, fetch skills, and combine multiple skills for a specific task.
The stack includes a Registry built on Hono and Drizzle on Postgres/pgvector, with a Next.js 15 site and a TypeScript CLI (@amitte-ai/cli). An Apache-2.0 licensed open model powers a small RAG assistant called Folio, which answers queries about where to find specific items. The project is still in its early stages, with plans to implement cryptographic signing for enhanced integrity checks. The project is open-source and welcomes community contributions and feedback.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
