Architect-Led AI Development — Automation Bundle

Reusable structure for AI-assisted development, based on the architect-led methodology.

---

What's in this bundle

/
├── README.md                             ← you are here
├── PROCESS.md                            ← the methodology (read this first)
├── architect-process-master.excalidraw   ← one-page visual overview
│
├── prompts/                              ← five prompt templates
│   ├── 01-frame.md
│   ├── 02-design.md
│   ├── 03-validate.md
│   ├── 04-implement.md
│   └── 05-review.md
│
└── skeleton/                             ← copy this into new projects
    ├── PROCESS.md                        (same as above, carried into every project)
    ├── README.md                         (explains the skeleton layout)
    ├── prompts/                          (same prompts, carried forward)
    ├── checklists/
    │   ├── before-coding.md
    │   ├── before-shipping.md
    │   └── regression-check.md
    ├── docs/
    │   ├── 01-frame.md                   (stub to fill in Phase 1)
    │   ├── 02-design.md                  (stub to fill in Phase 2)
    │   └── scenarios.md                  (stub to fill in Phase 3)
    ├── build/
    │   ├── build.js                      (working stub)
    │   └── validate.js                   (working stub)
    ├── authoring/                        (empty — humans edit here)
    ├── runtime/                          (empty — application code here)
    └── dist/                             (empty — build outputs here)

How to use this on a new project

1. Read PROCESS.md once. Ten minutes. Don't skip it.
2. Open architect-process-master.excalidraw at excalidraw.com (File → Open) to see the whole process on one page.
3. Copy skeleton/ into your new project repo. Rename it to whatever the project is called.
4. Open a fresh Claude conversation. Work through prompts/01-frame.md using your feature request. Save Claude's output to docs/01-frame.md.
5. Continue through Phases 2–5 the same way. Each prompt blocks Claude from regressing to code-first mode until the phase is complete.
6. Print the checklists and tape them to your monitor until the process is reflex.

What makes this different

Most "AI coding best practices" are lists of tips. This bundle is an enforceable process:

• The prompts refuse code output in phases 1–3. If you catch yourself accepting code early, you know you've regressed.
• The validator is the first code written. This is the biggest single ROI step — it catches design flaws before any runtime exists.
• The checklists are decision gates, not advice. Every gate either passes or fails; no middle state.
• The skeleton forces three-layer separation (authoring / build / runtime) — you can't accidentally collapse them.

When NOT to use this

Throwaway scripts under 100 lines. One-shot data transformations. The overhead isn't worth it.

The test: will anyone (including future-you) read this code after the task is done? If yes, use the process. If no, just ask the AI directly.

Origin story

This bundle came out of a PoC project — a chess opening trainer that was attempted seven times with vibe-coding and failed each time. On the eighth attempt, with the architect-led process applied, it worked on the first run. See the blog post for the full story if you want the narrative; this bundle is what distills the lesson into something reusable.

Maintenance

Update this bundle when a project teaches you something that extends the methodology. The point is not to have a perfect doc — it's to have a current doc that reflects what actually works for your team.

If the process ever becomes a checkbox exercise that doesn't produce better software, it has become the thing it was designed to replace. Retire or reform it then.