World Building for AI

Build the World before the product. This is the hands-off but non-generic path:

1. Context build
2. Synthesis
3. Build

Use this skill to turn a short prompt into a research system with explicit pillars, a tailored folder tree, parallel subagents, layered synthesis, critical review, and only then downstream specs or implementation plans.

Start Here

Read these bundled references before doing substantial work:

1. references/mode-selection.md — choose the correct depth
2. references/pillars.md — map the smallest complete set of pillars
3. references/folder-structure.md — create the right scaffold
4. references/orchestration.md — run the research phases correctly
5. references/five-sentence-superprompt.md — use when the user wants the shortest shareable version

Modes

Choose the lightest mode that still gives the user the right decisions:

• foundational — first-principles understanding and domain grounding
• deep-research — broad evidence gathering before a build
• world-building — the full harness: pillars, folders, research, synthesis, critique, and evidence-linked downstream docs

If the user says "build the world first", "deep research before building", or explicitly asks for a harness, default to world-building.

Workflow

1. Restate the build target in one sentence. Infer low-risk missing details instead of stopping.
2. Choose the mode.
3. Create the scaffold:
   • 00-brief/vision.md
   • 00-brief/world-pillars.md
   • 00-brief/success-criteria.md
   • tailored 01-research/, 02-synthesis/, 03-critical-review/, 04-product/, and 05-build/
   • WORLD.md as the living source of truth
4. Define pillars:
   • purpose
   • priority
   • key questions
   • dependencies
   • output paths
5. Launch research:
   • run one lead researcher per pillar
   • use parallel subagents where the environment supports them
   • keep raw evidence in 01-research/
6. Synthesize in layers:
   • per-pillar synthesis in 02-synthesis/per-tier/
   • cross-synthesis, evidence map, and decision framework
   • critical review for contradictions, bias, implementation reality, missing voices, and failure patterns
7. Derive downstream deliverables only after synthesis:
   • PRD
   • UX flows
   • acceptance criteria
   • architecture
   • backlog
   • roadmap
   • strategy docs
8. Iterate:
   • update WORLD.md first
   • then synthesis
   • then product docs
   • then code or execution plans

Operating Rules

• Prefer breadth first. Narrow only after evidence exists.
• Use current sources when facts are unstable.
• Mark assumptions, contradictions, and uncertainty explicitly.
• Separate raw evidence, synthesis, critique, and decisions.
• Treat marketing claims as untrusted until corroborated.
• Include contrary evidence and failure cases, not just success stories.
• If Task is unavailable, preserve the same phases sequentially instead of skipping them.
• Tailor the folder tree and pillar set to the request. This is a system, not a rigid checklist.

Research File Contract

Use this structure for each lead research file unless the task clearly needs a variant:

# [Topic] Research Report

## Executive Summary
- Key finding
- Confidence
- Product or strategy implications
- Research gaps

## Detailed Findings

## Sub-Agent Requests

## Design Implications

## Sources

Minimum Deliverables

• WORLD.md
• 00-brief/world-pillars.md
• structured 01-research/ outputs for each active pillar
• 02-synthesis/cross-synthesis.md
• 02-synthesis/evidence-map.md
• 02-synthesis/decision-framework.md
• at least one evidence-linked downstream deliverable in 04-product/

Trigger Phrases

Treat all of these as strong signals to use this skill:

• "deep research"
• "foundational research"
• "world building"
• "deep research before building"
• "build the world first"
• "research the whole space before we code"
• "hands-off but non-generic"