eddie

EDDIE

Explore → Define → Design → Implement → Evaluate.

A meta-skill plugin for Claude Code that walks non-technical builders through the same disciplined intent → design → create flow a professional product team would use. Research is a first-class citizen. Decisions are recorded with their why. Works for software, craft projects, process redesigns, and research write-ups.

EDDIE is a pushback engine, not a judge. Every gate makes you think harder. The user always controls whether to proceed.

---

Why this exists

I'm a non-IT founder building in public. I'd been using superpowers:brainstorming to design features for my site, and two failure modes kept biting me:

1. No durable record of why. The brainstorm gave easy multiple-choice answers and moved on. Six weeks later, when I went to extend something, nobody — including me — could remember why we picked B over A. There was no anchor doc.
2. The AI defaulted to "what it knew" rather than "what was actually most stable, efficient, or domain-correct." Without forced research, the agent would happily pick the obvious-but-not-best option, and being non-technical, I had no way to push back.

EDDIE fixes both: forced domain research before architecture, and durable PRD + architecture documents that survive 3-month gaps.

A third thing emerged while designing it: this discipline isn't software-specific. So EDDIE also runs for robotics build guides, process redesigns, and research write-ups — same five phases, adapted artifacts.

Who it's for

• Non-technical founders, hobbyists, and process designers who want professional-grade output without a CS degree.
• Engineers who keep losing the thread between "we decided X" and "why did we decide X."
• Anyone who's been burned by AI-generated code that looked fine and broke six weeks later.

If you've used superpowers:brainstorming and felt the gap, EDDIE is the opinionated version that closes it.

Install

System-wide:

claude plugin install https://github.com/m3m0ng/eddie

Project-local (for development or one-off use):

claude plugin install --plugin-dir ./eddie

Usage

In any project:

/eddie

The orchestrator checks whether there's an active EDDIE run, classifies the project type (or resumes a prior run), and walks you through the active phases one at a time. Hard gates between every phase — you can always revise or stop.

What you get

For each EDDIE run, a self-contained folder under <project>/eddie/<run-slug>/:

• interview.md — full Explore + scope synthesis with your verbatim quotes
• prd.md — product requirements with Given-When-Then acceptance criteria
• architecture-design.md (or approach.md for non-software) — the why behind every major decision, ADR-style
• tasks.md — vertical-slice task list keyed to PRD requirements
• evaluation/rtm.md — Requirements Traceability Matrix mapping each PRD requirement to the test that proves it
• .eddie-config.json — run state for resumption

Multiple runs per project coexist. Tests accumulate at <project>/tests/<run-slug>/. The full project test suite re-runs at the end of every EDDIE round as a regression check, so adding a feature can never silently break a previous one.

Phases at a glance

Phase	What it does	Output
Explore	Two-pass interrogation: raw-intent first pass → market-research subagent → evidence-armed second pass. Build-vs-existing-alternatives is informational, not adversarial.	interview.md + research-findings.md
Define	First-draft PRD, then probe edge cases, anti-patterns, YAGNI.	prd.md
Design	Parallel research subagents on existing solutions + technical feasibility. ADR-style architecture decisions.	architecture-design.md or approach.md
Implement	Vertical-slice task list. TDD enforced via Red/Green/Refactor subagent isolation.	tasks.md + working code + tests
Evaluate	Testing Trophy default. Continuous companion to Implement (per-slice tests) + final wrap-up (E2E + LLM-judge).	evaluation/rtm.md + full project test suite

Adaptive by project type

The orchestrator classifies each run as one of: software-app, software-script, craft-physical, process-redesign, research-doc, hybrid. Phase activation adapts:

• software-script skips Design unless requested.
• craft-physical swaps architecture-design.md for the lighter approach.md, replaces TDD with step-by-step guides, and uses a human-observation rubric for Evaluate.
• process-redesign skips Implement (the redesigned SOP itself is the deliverable).
• research-doc runs Explore + Define + Design only.

Classification is per-run. A single project can host a software-app run today and a process-redesign run next month.

Design principles