Manifest
The manifest is the first pillar of Lytos: Intent.
It answers the question every AI agent needs before doing anything useful: what is this project, and why does it exist?
What it contains
Section titled “What it contains”The manifest is a single file — manifest.md — placed at the root of .lytos/. It defines:
| Section | What it answers |
|---|---|
| Identity | Name, description, owner, repo, version |
| Why this project exists | The problem it solves, the value it creates |
| What it is / is not | Clear boundaries to prevent scope creep |
| Tech stack | Language, framework, database, CI/CD, hosting |
| Project vocabulary | Domain terms the AI must understand |
| Fundamental constraints | Non-negotiable rules (offline-first, zero deps, no telemetry…) |
| Development principles | How to decide when two approaches seem equal |
Why it matters
Section titled “Why it matters”“Delegating to AI without structure is outsourcing without a brief.”
Without a manifest, an AI agent produces generic code. It doesn’t know if your project is a library or a SaaS, if you prefer REST or GraphQL, if you deploy to AWS or a Raspberry Pi.
The manifest is read at the start of every session. It’s the first thing the agent loads, before skills, rules, or memory.
Example
Section titled “Example”## Identity
| Field | Value ||-------|-------|| Name | bookshelf-api || Description | REST API for managing book collections || Owner | Jane Doe (@janedoe) || Stack | Python 3.12, FastAPI, PostgreSQL, Docker |
## Why this project exists
Libraries need a modern API to manage their collections.The existing system is a 15-year-old PHP monolith.
## Fundamental constraints
- All endpoints must return JSON- No breaking changes without a deprecation period- Response time < 200ms for read operationsBest practices
Section titled “Best practices”- Be specific, not aspirational. Write what the project is, not what you wish it were.
- Update it when the project evolves. The manifest is a living document.
- Keep it under 150 lines. If it’s longer, some content belongs in memory or rules.
Learn more
Section titled “Learn more”- How to write a CLAUDE.md that works — what to put in the file the AI reads first, what to leave out, and how to keep it maintainable.
- AGENTS.md vs CLAUDE.md vs cursor rules: which to use — how the three major conventions relate, and why you probably don’t have to choose.