The Process
Five phases.
One session.
From PRD to deployed product. Every engagement follows the same pipeline. The state machine enforces it. Marcus makes sure it doesn't skip steps.
Debate
Tension before consensus.Rounds 1–2
Steve and Elon read the PRD independently and form positions across all six deliverable areas: product design, brand, customer personas, marketing, go-to-market, and engineering. They stake these positions without coordination. Then they challenge each other. No deference. The point is to surface every real disagreement before a line of code is written.
Outputs
- →Independent position statements per deliverable area
- →Challenge rounds logged in rounds/
- →Strategic disagreements surfaced and argued
Why this matters
Agencies that agree too quickly ship mediocrity. Real disagreement between high-standard collaborators produces decisions that survive contact with reality.
Plan
Decisions lock. Teams get hired.Round 3
Directors converge. Marcus mediates any unresolved conflicts and formally logs the locked decisions. Steve defines his sub-agent team (designer, copywriter, brand strategist). Elon defines his (market analyst, growth strategist, engineer). Each agent gets specific inputs, outputs, and a quality bar they can't dodge.
Outputs
- →Locked decisions document
- →Agent definitions in team/
- →Sub-agent assignments with clear inputs and quality gates
Why this matters
Ambiguous mandates produce ambiguous output. Every agent knows exactly what they own and what done looks like.
Build
Parallel execution. No hand-holding.Rounds 4–8
Sub-agents execute their assignments in parallel. Each reads the PRD, the locked decisions, and their role definition. Output goes to deliverables/{project}/drafts/. Directors can intervene if output drifts from strategy. Blockers surface in STATUS.md — the agency doesn't silently fail.
Outputs
- →Design specifications
- →Codebase with tests
- →Copy, messaging, brand guide
- →Engineering specs and architecture
Why this matters
The factory is the product. How things get built — in parallel, with clear ownership, against stated quality bars — is what makes the output trustworthy.
Review
Two directors. One standard.Round 9
Steve audits every deliverable for taste, craft, and brand consistency. Elon audits for feasibility, accuracy, and market alignment. Jensen provides board-level strategic feedback. Marcus checks consistency across the full set. Revisions go back to specific agents with precise feedback — not vague notes.
Outputs
- →Revision requests with specific agent assignments
- →Approved deliverables promoted to final/
- →Jensen's strategic notes
Why this matters
Quality gates are only meaningful when they have teeth. Nothing ships that Steve wouldn't show to someone he respects and Elon can't justify from first principles.
Ship
Production. Memory. Done.Round 10
Final deliverables assembled in deliverables/{project}/final/. Production deployment initiated. Steve and Elon write a joint summary. Marcus logs learnings to memory/. The organizer consolidates and prunes. STATUS.md returns to idle. The agency is smarter than it was before.
Outputs
- →Production deployment
- →Joint executive summary
- →Memory updates for future projects
- →STATUS.md → idle
Why this matters
The agency compounds. Every project teaches it something. The next PRD gets better output because this one happened.
The Edge
What makes it different.
Most teams agree too quickly. Steve and Elon don't.
Consensus is the enemy of excellence. The debate phase exists because real disagreement between thoughtful people produces decisions that survive first-principles scrutiny. You pay for conflict. What comes out the other side is defensible.
Standards
What we will not compromise on.
No mediocrity.
If it's not worth building beautifully, it's not worth building. Good enough is the enemy of insanely great.
No hand-waving.
Every claim survives first-principles scrutiny. Beauty must be defensible. Feasibility must be proven.
No feature bloat.
Focus means saying no to good ideas so great ideas can breathe. The agency kills more features than it ships.
No slow timelines.
Timelines are compressible. Physics is the only constraint. Default to faster until the physics say otherwise.
No silent failure.
Blockers surface immediately in STATUS.md. The agency doesn't pretend it isn't stuck. It flags, pivots, and escalates.
No forgetting.
Every project updates memory. The organizer consolidates. The agency compounds. Client three gets the benefit of clients one and two.
When things go wrong
The retry protocol
1. Agent fails → retry once with the same prompt
2. Retry fails → try an alternative approach
3. Alternative fails → mark as blocked in STATUS.md
4. 3 failures on the same task → stop, engage human
5. While blocked on one task → work on unblocked tasks
See it in practice
The LocalGenius case study walks through all five phases — the debate transcripts, the build output, the final numbers.