Kelly Tweets: BMAD Methodology

Summary: BMAD (Build My Idea Automated Development) is a Gauntlet AI framework for orchestrating AI agents to build software from a single prompt. Kelly experimented with BMAD extensively, initially running hybrid multi-model pipelines (GPT for ideation → Opus for PRD → Gemini for designs → Codex for infra → BMAD for app build → Codex for security review) before concluding that the cross-model handoffs created too much noise. The verdict: Kelly's own factory outperformed hybrid approaches.

Key Concepts

• **BMAD:** Build My Idea Automated Development — a Gauntlet AI framework for idea-to-app AI orchestration
• **Wiggum:** Related Gauntlet AI orchestration framework, used alongside BMAD
• **Hybrid pipeline failure:** Passing work between multiple models (GPT → Opus → Gemini → Codex → BMAD) lost context and broke the output
• **Factory > Hybrid:** Kelly's own factory (single coherent orchestration) beat the hybrid approach

Notable Patterns

The Failed Hybrid Experiment

Kelly ran this pipeline:

1. GPT for idea generation + analysis
2. GPT ↔ Opus back and forth for idea → PRD
3. Gemini for PRD → wireframe → designs
4. Codex builds infra (stack pre-selected)
5. Bastardized BMAD/Wiggum to build app
6. Codex security review

Result: Total failure. A council of models graded the output — Claude won universally. Too much noise in the model handoffs.

"Too much noise happened in the passing between models. Too cute."

The Working Alternative

Kelly's own factory — a single coherent orchestration with hard rules, pass/fail guardrails, and scripted sub-agent communication — outperformed all hybrid approaches.

Gauntlet BMAD Improvements

Gauntlet AI grads worked on tweaks of BMAD/Wiggum to push how far autonomous build chains can go. The approach: start with the same PRD, run it through various models, orchestration methods, progression stacks to identify where it goes off the rails, tweak, try again.

"Like bug squashing except treating all of AI software development like a production line."

Second Experiment: 2-3 Model Hybrid

Working with 2 or 3 models at their greatest strengths with maxed context windows allowed passing more chains of commands without losing as much context.

Example: Codex infra → Claude build → Codex security review

The Real Pattern

"Every 1% less you need humans in the loop results in 10x speed-up in building software."

BMAD showed that the bottleneck isn't the AI's capability — it's the orchestration layer between models and the human loop.

Kelly's BMAD-Derived Factory

Rather than using BMAD as a black box, Kelly evolved it into her own factory with:

• Shell scripts as quality gates (enforcement layer AI agents can't bypass)
• Hard rules that must be followed before progression
• Consistent sub-agent communication patterns
• Retry logic that self-improves

Related

[[kelly-tweets-factory]], [[kelly-handbook-software-factory]], [[austen-relevant]]