How BeforeTomorrow Works
BeforeTomorrow generates world-class presentations by orchestrating 15+ specialized AI agents (Claude Opus 4.6 + Sonnet 4.6 on AWS Bedrock) through a dependency-aware parallel pipeline. Each deck passes a 13-critic quality gate with automated Reflexion repair. The system uses Tree-of-Thoughts planning with 8 narrative structures, producing unique per-slide design compositions — no templates — in 10-15 minutes per artifact.
What is the generation pipeline?
The pipeline follows 6 phases executed by the Council Orchestrator. Each phase uses dependency-aware parallel execution to minimize wall-clock time while maximizing quality.
| Phase | Agents | Output |
|---|---|---|
| 1. Planning | Tree-of-Thoughts Planner (Opus 4.6, 8 branches) | Narrative structure + slide budget |
| 2. Research Wave | Outline ∥ Domain ∥ Researcher ∥ Storyteller ∥ Investor-Psychology (parallel) | Structured intelligence from 5 perspectives |
| 3. Validation Wave | Brand ∥ Financial-Auditor ∥ Fact-Checker (parallel) | Brand compliance + financial accuracy + fact verification |
| 4. Creative Wave | Visual Director ∥ Copy Director ∥ Chart Director (parallel) | Per-slide design + professional copy + data visualization |
| 5. Post-Production | Copy-Editor ∥ Asset-Prompt-Craftsman ∥ Design-Architect (parallel) | Polish + asset prompts + unique layouts per slide |
| 6. Quality Gate | 13 critics + Synthesis Orchestrator + Reflexion repair (up to 2 passes) | Final deck — ship only if quality threshold met |
What are the key performance numbers?
| Metric | Value |
|---|---|
| Total agents per deck | 15+ ULTRA-grade |
| Quality critics | 13 (narrative, design, fact, constitutional, competitive, coherence) |
| Agent prompt length | 4,000-7,000 words each |
| Narrative structures | 8 (including three-act, pyramid, inductive, problem-solution) |
| Layout compositions | 32 canonical entries |
| Generation time | 10-15 minutes (was 21 min, optimized via targeted repair) |
| Cost per deck | $3-20 LLM compute (no cap) |
| Industries supported | 140+ with domain-specific intelligence |
| AI models | Claude Opus 4.6 (creative) + Sonnet 4.6 (extraction) |
| Data residency | EU (AWS Bedrock cross-region profiles) |
What is domain engineering?
Domain engineering means each studio models a real business domain — not just a prompt template. The Slide Studio doesn't ask an LLM to "make slides." It models how a world-class creative agency operates: a research team gathers intelligence, a storyteller crafts narrative, a visual director designs each slide uniquely, a copy director writes at the highest professional level, and critics review everything before delivery.
Each agent has a 4,000-7,000 word ULTRA system prompt defining their role profile, frameworks they own, day-to-day responsibilities, output rubric, anti-patterns, mandatory output contract, and self-verification checklist. This is domain engineering — not prompt engineering.
How does the 13-critic quality gate work?
After the creative pipeline completes, every deck enters a quality gate with 13 specialized critics running in parallel. Critics evaluate narrative coherence, design courage (anti-carbon-copy), competitive differentiation, factual accuracy, constitutional compliance, adversarial edge cases, and more. A Synthesis Orchestrator (Opus 4.6) arbitrates — if quality threshold is met, the deck ships. If not, targeted repair revises only the flagged slides (not the whole deck) and re-runs only the failed critics. Up to 2 repair passes ensure every deck meets the quality bar.
How does real-time fact verification work?
The Fact Checker agent uses a ReAct + CRITIC methodology: it extracts up to 24 claims from research data, verifies each against tier-1 sources (academic, government, industry) via web search, and classifies claims as verified, contradicted, unverified, or skipped. This evidence pack is injected into the LLM evaluation — so fact-checking is grounded in real evidence, not training-data hallucination. Based on Gou et al. 2023 CRITIC (arXiv:2305.11738), this reduces hallucination rates by approximately 30%.