Eidos

Eidos

Spec-driven development for Claude Code.

Markdown specs capture intent.
Code is a downstream manifestation of that intent.
The two stay in sync bidirectionally: specs shape code, code can update specs, and conflicts become human decisions.

Why

The hard part of AI-assisted development is not writing code.
It is keeping human and AI on the same page.

Intent lives in ephemeral places.
The chat that scrolls away. A commit message. Your head.
The AI acts on a fraction of it, and each new session starts blind.

Eidos has one goal.
Human and agent hold the same picture of the work, at every point.
It gets there three ways:

- Capture intent as artifacts. Specs, claims, plans, decisions. Structured files, not disappearing chat.
- Keep the human steering. Surface reasoning before acting, checkpoint at decisions, so course-correction is cheap and early.
- Persist and reload. Those artifacts load back into session context automatically, so understanding compounds instead of resetting.

Specs are the flagship.
They describe what a system *should* be, and stay in sync with the code.
When the two drift, neither automatically wins. The human decides which to update.

Three folders

Specs describe timeless intent.
Plans describe time-bound work.
Code is the result.

With the agent-inbox feature enabled, a fourth folder appears: agent-inbox/.
Each repo becomes an agent that coordinates with peer repos through append-only messages there.
It is off by default. Turn it on with /eidos:config.

The core loop: spec ↔ code

Green-field

Write a spec for what you want: behaviour, design, verification.
Do it by hand, or let /eidos:spec walk you through it.

Then push it into code:

Verify with tests, by hand, or both.
If something is off, update the spec, commit, and push…