Eidos
Spec-driven development plugin for Claude Code.
Markdown specs capture intent.
Code is a downstream manifestation of spec intent.
The relationship is bidirectional: specs shape code, code can update specs, and conflicts become human decisions.
Why
Development intent lives in chat messages, commit messages, and the developer's head.
When an AI starts a new session, it sees code but not the reasoning behind it.
Eidos keeps intent in spec files alongside the code.
Specs describe what a system should be.
When spec and code drift apart, neither automatically wins — the human decides which to update.
Three folders
project/
eidos/ # what it SHOULD be (specs, claims)
memory/ # how we got here (plans, decisions, sessions)
src/ # what it IS (code)
Specs describe timeless intent.
Plans describe time-bound work.
Code is the result.
How it works
Green-field
Write a spec to capture what you want — behaviour, design, verification.
This can be manual or collaborative (/eidos:spec walks you through it).
Then push it into code:
/eidos:push my-feature-spec
Verify — with tests, manually, or both.
If something's off, update the spec, commit, push again.
/eidos:push recent 1 picks up the latest spec change and its precise diff.
Repeat.
Existing project
You already have the manifestation — mine it for intent:
/eidos:pull the auth module
Review the extracted spec, refine it, and iterate from there.
Now you have a spec to push against, drift-check, and evolve.
Plans and memory
Multi-step work gets a plan:
/eidos:plan migrate the database layer
Plans live in memory/, persist across sessions, and track progress phase by phase.
Resume with /eidos:plan-continue.
Other procedural skills accumulate context the same way:
/eidos:research — investigate external topics, document findings with sources/eidos:observe — capture testing issues mid-plan, update specs, inject tasks/eidos:decision — record architectural choices with options and rationale
Externalise everything
AI "learning on the job" is just capturing ephemeral knowledge into files and composing them into session context.
That knowledge can be anywhere — your chat history, a coworker's head, legacy code nobody dares touch.
Eidos gives it all a place to land and a way to get loaded when it matters.
Install
Requires Claude Code.
# Add marketplace
/plugin marketplace add agenticnotetaking/eidos
# Install plugin
/plugin install eidos@eidos
Then restart Claude and run /eidos:init inside any project to bootstrap the folder structure.
Alternative: clone and alias
If you'd rather run from a local checkout (easier to adjust eidos to your needs), clone the repo and add shell aliases:
git clone https://github.com/agenticnotetaking/eidos.git ~/repos/eidos
Then in your ~/.bashrc or ~/.zshrc:
export EIDOS_DIR="$HOME/repos/eidos"
eidos() {
claude --plugin-dir "$EIDOS_DIR" "$@"
}
eidos-auto() {
claude --dangerously-skip-permissions --plugin-dir "$EIDOS_DIR" "$@"
}
Adjust the path if you cloned elsewhere. eidos-auto skips permission prompts — only use it in directories you trust.
Skills
Run /eidos:help for the full list. Highlights by category:
Core loop — sync between spec and code:
| Skill | What it does |
|---|
/eidos:spec | Create a spec via structured Q&A |
/eidos:push | Implement code to match a spec |
/eidos:pull | Reverse-engineer a spec from existing code |
/eidos:drift | Analyse divergence between specs and code |
/eidos:sync | Bidirectional reconciliation |
/eidos:coherence | Check specs against each other for contradictions |
/eidos:weave | Discover missing wiki links, prune stale ones |
Planning — structured multi-step work:
| Skill | What it does |
|---|
/eidos:plan | Plan multi-step work with phases and actions |
/eidos:plan-continue | Resume work on an existing plan |
/eidos:research | Research a topic and document findings |
/eidos:decision | Record a decision with options and rationale |
/eidos:brainstorm | Explore ideas around a topic |
/eidos:experiment | Log-based iterative exploration |
Observation — understand what you have:
