Authoring tools · experimental

You know what you want. Getting an agent to build it is the hard part.

SeedSpec helps you write down what you actually know about a product — the outcome, the boundaries, the decisions that have no obviously right answer — so an agent builds that instead of something plausible.

Why bother writing it down

An agent asked to build something will make every decision you left out. It will make them confidently, and it will not tell you which ones it made. Usually they are fine. The ones that are not are expensive, and you find them later.

Start from a seed

Write what you want to exist in plain language. Your agent reflects it back, asks about the decisions only you can make, and records your answers where an implementer will actually see them.

Keep what only you know

The choices that make a product yours — what happens on a failed check, who can approve what, where the money goes — are the ones an agent invents when nobody wrote them down. Authoring is how you stop guessing from becoming policy.

Hand it to any agent

The result is ordinary files with a version and a digest. Give it to Claude, Codex, or whatever you use next year. Nothing about the tool you authored in follows it.

What authoring actually looks like

01

Describe what you want

Three paragraphs is a real starting point. You are not writing requirements and you are not designing the implementation.

02

Answer what you are asked

Questions come one at a time, and every one is triggered by something you already wrote. Nothing is added because similar products usually have it.

03

Say what success looks like

Results someone could observe, not a test plan. This is what an implementing agent checks itself against.

04

Share it

A versioned package anyone can hand to their own agent, in their own environment, with their own connectors and their own security.

It stays a conversation. Nothing changes in your package until you say so, and declining a suggestion leaves nothing behind.

Try it now

You need Node and an agent you already use. There is nothing to sign up for.

npx @seedspec/cli init application --output my-package cd my-package npx @seedspec/cli author

Or start from an existing package instead of a blank page:

npx @seedspec/cli begin https://github.com/SeedSpec/reference-solutions/tree/main/solutions/family-hub/seedspec

begin validates the package and shows you what it contains. It does not run anything the package carries.

Underneath

What you author is a SeedSpec package: ordinary Markdown and a small manifest, with a versioned protocol behind it so the meaning survives moving between people, agents, and projects. You never have to look at that layer, but it is fully specified and open if you want to.

Read the protocol at seedspec.dev →