| name | gemini-vault-fusion |
| description | After completing reverse engineering vault, use a high-reasoning LLM (Gemini 3.1 Pro Preview or similar) to review all documentation, identify inconsistencies, generate a professional PRD/Blueprint structure, then FUSE the AI output with existing deep technical docs into a v3.0 definitive document.
|
| tags | ["reverse-engineering","document-review","gemini","ai-assisted","fusion-strategy","prd","blueprint"] |
Gemini Vault Fusion — AI-Assisted Documentation Review & Consolidation
When to Use
After completing a large reverse engineering vault (30+ documents, 20+ modules, ~200KB+), when you need to:
- Audit for inconsistencies across documents
- Generate a professional-quality PRD/Blueprint with personas, journeys, acceptance criteria, metrics
- Not lose technical depth in the process
The Problem
- AI-only output (feeding vault to Gemini and asking for PRD): produces professional structure but loses deep technical content — schemas get simplified, UX divergences disappear, state machines become generic, risks/implementation waves vanish.
- Human-only docs: have glorious technical depth (797-line schemas, 7 UX divergences, 4 state machines) but lack executive polish — no formal personas, no acceptance criteria section, no success metrics.
The solution is not to choose. It's to FUSE.
The Fusion Strategy
Phase 1 — Vault Consolidation
- Gather ALL documentation files from the project
- Consolidate into ONE file with clear source headers:
## Fonte: docs/caminho/arquivo.md
[full content of file]
- Save to
/tmp/vault_completo.md
Phase 2 — AI Review (Gemini 3.1 Pro Preview)
Feed the consolidated vault to a high-reasoning model with a structured audit prompt:
Revise toda esta documentação de engenharia reversa do [SISTEMA].
Identifique:
1. Inconsistências entre documentos (com documento e linha)
2. Informações faltantes
3. Contradições
4. Sobreposições
5. No final: nota de 0-10 para completude e consistência
Save output as docs/AUDITORIA-GEMINI-[MODELO].md
Phase 3 — AI Generation (Second Prompt)
Com base em toda esta documentação, produza:
1. PRD completo e profissional — visão geral, personas, funcionalidades,
regras de negócio, jornadas de usuário, critérios de aceitação
2. Blueprint arquitetural — stack, schema, multi-tenancy, API design,
UI architecture, infraestrutura
Save outputs as draft v2.0.
Phase 4 — Human Cross-Validation
Deploy a dedicated auditor ({{AUDITOR}} role) to compare AI output vs existing docs:
| Aspect | AI v2.0 | Human v1.x | Veredict |
|---|
| Structure/Personas | ✅ Professional | ⚠️ Lacking | 🏆 AI wins |
| Deep schemas | ❌ 15 generic | ✅ 30+ detailed | 🏆 Human wins |
| UX divergences | ❌ Zero | ✅ 7 documented | 🏆 Human wins |
| State machines | ⚠️ Incomplete | ✅ Complete | 🏆 Human wins |
Phase 5 — Fusion (v3.0)
Strategy: Use AI v2.0 as the shell (executive structure, personas, journeys, acceptance criteria, metrics), then inject the deep technical content from human v1.x (schemas, UX divergences, performance targets, state machines, risks, implementation waves).
Result: PRD-v3.0.md and BLUEPRINT-v3.0.md — professional AND implementable.
Phase 6 — Fix Real Gaps Found
The AI audit will find real gaps. Fix them:
| Priority | Gap | Action |
|---|
| 🔴 | Schema contradictions | Update source docs immediately |
| 🟡 | Architecture gaps (Docker networking, auth) | Document resolution |
| 🟢 | Bloated files | Clean up |
⚠️ Pitfalls
-
Model names change fast. Always verify available models via API before stating what exists. {{COMMANDER}} corrected us that Gemini 3.5 Flash and 3.1 Pro Preview existed when we thought they didn't.
from google import genai
client = genai.Client(api_key=key)
for m in client.models.list():
print(f'{m.name}: {m.supported_actions}')
Don't guess model names — check the API.
-
AI loses technical depth. The v2.0 will simplify schemas, drop UX divergences, and genericize risks. Plan the fusion BEFORE generating, not after.
-
The shell is the easy part. 80% of the v3.0 fusion effort is identifying what the AI dropped and re-injecting it. Budget time for this.
-
Cross-validation is essential. Never trust AI-generated documentation without a human auditor comparing it point-by-point against the originals.
-
Context window management. A vault of 193KB+ fits in Gemini 2M context, but the response may truncate. Split into chunks if needed, or use a model with large context (Gemini 3.1 Pro Preview = 2M tokens).
Model Selection Guide
| For | Model | Why |
|---|
| Deep review + generation | gemini-3.1-pro-preview | Best reasoning, 2M context, catches inconsistencies |
| Fast iteration / re-generation | gemini-3.5-flash | Faster, cheaper, still capable |
| Vision (screenshots) | gemini-2.5-flash (legacy) or Kimi K2.5 via opencode-go | For image analysis during RE |
Example Output
This skill was developed and validated on the {{PROJECT_NAME}}/Dontus RE project:
- Input: 18 documents, 28 modules, 193KB vault
- AI Model:
gemini-3.1-pro-preview
- Audit found: 5 real inconsistencies (Vue.js vs Alpine.js, FK direct vs N:N, Docker networking missing, auth cross-tenant undocumented, bloated cadastros doc)
- Fusion result: PRD v3.0 (7.7KB) + Blueprint v3.0 (7.2KB)
- Time: ~3 hours total (1h vault consolidation + 30min AI generation + 1h cross-validation + 30min fusion)
Related Skills
re-audit-consolidation — For the pre-fusion audit phase (multi-agent coverage assessment)prd-clone-exhaustivo — For the initial mapeamento/discovery phase
