AI Engineering

Product Development Framework

A Claude Code plugin that compiles fuzzy product ideas into validated, agent-ready spec packages — the guessing engineered out of the handoff.

Role · Designer and builder

Outcomes

  • In production across my own client work — the same prompt chain scoped the grounded RAG chatbot engagement
  • Nine ADRs (0001–0009) document every architecture decision, including the one that killed the MCP server
  • Public and installable via the Clownware plugin marketplace

Stack

  • Claude plugins
  • Skills
  • Subagents
  • Prompt systems
  • ADRs

Context

A Claude Code plugin that walks a fuzzy product idea through the full early funnel — problem definition, personas, hypothesis, user flows, screens, technical spec — and compiles the answers into a spec package an implementation agent can build from. Shipped publicly as a Clownware product.

The Problem

Implementation agents don’t fail because they can’t code — they fail because they guess. AI prototyping without UX rigor produces plausible junk; UX process without AI is slow. Neither failure mode is acceptable when the prototype is the thing that decides whether a project gets funded.

My Role

Designer and builder — the process encoded is the one I run on real client engagements.

Approach

  • Encode the rigor into the workflow. Each phase is a skill with defined inputs, outputs, and quality gates; subagents handle the specialized passes.
  • The artifact is a spec package, not a chat transcript. A manifest, context (the why), specification (the what), and a docs layer recording what was decided — and what was explicitly excluded — topped with a CLAUDE.md handoff any implementation agent can pick up cold.
  • A package that doesn’t hold together doesn’t compile. The compile step cross-checks every entity, flow, screen, and endpoint reference across the spec files, so broken handoffs surface at spec time instead of build time.
  • One spec, any starter. The package is implementation-agnostic: the same spec hands off to the Astro or Go performance starter, and the implementing agent builds under that starter’s constitution. The spec says what to build; the starter enforces how.
  • Earn the complexity. The first architecture was an MCP server. It was premature infrastructure — a plugin was the right weight. The reversal is documented in the ADR log, not hidden.
  • Real-world validation. The same prompt chain scoped the grounded RAG chatbot engagement — the system is exercised by paying work, not demos.

The discipline running through this approach is the site’s thesis: Architecture Is the Interface Now.

What I’d Do Differently

Start with the plugin. The MCP detour cost a rewrite — a useful scar, and the reason “earn the complexity” is now a rule I apply before choosing any infrastructure.

All work