HTTP carries hypertext. SMTP carries mail. TCP carries bytes. Every protocol defines a capsule — a unit of transfer — and the semantics governing what happens when it arrives. Software built on large models has no such primitive. Context — the structured information a model needs to act — has no standard wire format. Skills, tools, and prompts ship as zip files, git repositories, bash scripts, or unstructured markdown. Nothing is interoperable. Nothing composes.

The bitter lesson of AI research is that scale beats craft. The same lesson is landing in software: context beats interfaces. A capable model does not need a perfectly-typed schema — it needs context that transfers reliably across boundaries. Seed/1.0 is the wire format for that transfer.

A seed is a single UTF-8 file. It is simultaneously valid markdown (the agent sees prose) and valid bash (the shell sees a self-contained install script). The dual-legibility invariant is the core constraint. Grammar below uses ABNF (RFC 5234).

The transport layer is POSIX sh. Line 1 of the seed MUST be a usage comment in the form specified by the grammar; it is the primary inspection surface in both reading contexts. The sentinel MUST be SEED_ followed by exactly 8 uppercase hexadecimal digits. The sentinel MUST be verified absent from the payload before emission; implementations choosing a colliding sentinel MUST regenerate. Backward-compatibility: the prefix XDOWN_ is a valid sentinel in conforming implementations of Seed/1.0.

The metadata layer is YAML 1.2 frontmatter. The fields seed and grow are MUST fields. Implementations SHOULD include type (one of skill, tool, context, data, archive) and name. Unknown fields MUST be preserved on round-trip and ignored during growing.

The payload is arbitrary UTF-8 content, typically markdown. It MUST NOT contain the sentinel. The dual-legibility invariant holds without transformation: to the agent, the heredoc delimiters read as a fenced block header and trailer; the payload is natural prose. To the shell, the heredoc writes the payload to disk unchanged.

When type: archive, the payload MUST conform to the Seed archive format: YAML frontmatter followed by HTML-comment file sections. The frontmatter MUST include marker: <6-hex>, root: <dir-name>, and at: <ISO-8601>. File sections use the delimiter <!--seed:MARKER@file path="..."-->; the backward-compatible form <!--fold:MARKER@...--> is also valid. Path safety rules from Section 7.2 apply.

Three operations are sufficient. They are composable and independently implementable. A conforming implementation of any one is not required to implement the others.

The grow: field is the receiver’s instruction set. It accepts a single directive (scalar) or an ordered YAML sequence. Execution is in listed order; fail closed on any failure.

Receivers MUST fail with a non-zero exit and a descriptive message on unknown directives. Silent failure is a protocol violation.

A conforming seed for a single-file skill. Transport (blue) wraps metadata (gold) wraps payload (green). This file is simultaneously paste-to-agent and curl | bash.

Any domain that implements the following endpoints is a conforming Seed registry. The registry stores seeds and serves them at stable URLs. Authentication for mutation operations uses opaque edit keys returned at plant time. The full registry specification is at /registry.

A conforming seed MUST: contain all three layers in fixed order; include seed: "1.0" and grow: in metadata; use a sentinel not present in the payload; begin with a usage comment on line 1.

A conforming pack implementation MUST produce conforming seeds. A conforming plant implementation MUST serve seeds at stable HTTPS URLs. A conforming grow implementation MUST resolve grow directives without silent failure.