An operating method for agentic builds
One small loop, run by any agent, carried across every session.
Seeds is a lightweight operating method for projects carried by agents across many sessions — built so every session ends in a verified result:
- Understand — context from the project itself, not chat history
- Decide — hard-to-reverse choices go to you; the rest proceed, recorded
- Slice — one bounded outcome at a time, with criteria you can actually check
- Build — that outcome and nothing else, in pieces you can check as you go
- Verify — proved on the real thing, not a simulation of it
- Hand off — record the evidence and where things stand, so the next session starts from reality
hand off plants the next understand — the loop replants itself
A small loop for high leverage: nothing lives only in chat history, done means it actually works, and any agent, any session, picks up where the last one stopped.
Start hereTry it on one piece of real work
1 · Hand your agent the method
Any agent, just paste either prompt:
Use Seeds (tryseeds.dev/agent) to set up this project.
Adopt Seeds (tryseeds.dev/agent) in this project. Reuse its existing conventions; add only what is missing.Or install the skill bundle (Claude Code, Codex, Cursor, and 70+ more):
npx skills add mauroibz/seeds2 · Kick off your first piece of work
Describe to your agent what you want to get done, and end with:
“Adopt Seeds in this project. Plan and execute, stop and report after it.”
The stakesWhat it establishes
Explicit sources of truth
No fact lives only in chat history. Every question has one authoritative answer: the plan owns scope, the records own history, the handoff owns what happens next.
One outcome at a time
A bounded piece of work with criteria you can check, and a clear line for what's not included. Small fixes take a light path instead of full ceremony.
Proof, not plausibility
A simulation of the thing being proved is not proof. Anything people can see or use gets exercised for real, on real data — because one product shipped broken behind 122 passing tests.
Durable history
Decisions get written down and never silently rewritten. A living handoff note states current reality and the exact next action for whoever — whatever — picks up next.
Agents work fast. What breaks projects is everything around the work: requirements that lived only in chat history, results that were never checked against reality, decisions nobody wrote down, and a session that died before the next one knew where to start.
Field notesWhy Seeds exists
Everything passed. The product did not work.
Seeds exists because of a day like that. In the project that led here, thirteen rounds of work produced an application with 122 passing backend tests, 38 passing frontend tests, 86% coverage — and it still failed its basic purpose. No required check had ever run the application and watched the flow.
The method was rebuilt so that can no longer happen quietly: nothing is "done" until it has been exercised for real, and the walkthrough gate exists to catch exactly this.
Go deeperThe paper trail
- GitHub repository — method, templates, and validator
- The method — the durable methodology
- FAQ — adoption questions, answered
- Changelog — versioned since 1.0.0
Seeds is tool-agnostic: no CLI, no required folder layout, no lock-in. Markdown files your agents read. CC BY 4.0.