🧬 Anatomy of a deliverable prompt
The 15 prompts aren't improvised—they all follow the same skeleton. Knowing this template is what lets you edit any of them and even create new deliverables without breaking the pipeline. Gemini receives the context (research + cheat sheets) and then the prompt with this structure.
// structure of a deliverable prompt (e.g., tech_inventory)
# Task: Generate Technology Inventory & Data Infrastructure Assessment Based on the company research and context provided above, create a comprehensive technology inventory document. ## Required Sections ### 1. Executive Summary - Current landscape, key gaps, AI-readiness (high-level) ### 2. Current Technology Stack | Category | Tool/Platform | Purpose | AI-Ready | Data Integration | |----------|--------------|---------|----------|------------------| ### 3. Data Infrastructure Assessment ... ### 5. AI Readiness Assessment (tabela 1-5 por dimensão) ### 6. Recommendations (0-30 / 30-90 / 90+ dias) ## Output Format - Use markdown; include tables; note assumptions; keep it actionable
💡 Why a table and why "above"
Tables force the model to structure information—and structure is what makes the document look like consulting work. “research and context above” is the hook: Phase 1 and the Trilha 2 cheat sheets come in before in the prompt, so the model writes about real facts, not in a vacuum.
the objective
the structure
the rules
comes back "above"
🔬 Evaluation: where the company stands
The first group diagnoses. It consists of three prompts that establish the factual basis of the entire package: without an honest assessment, the roadmap becomes a guess and the ROI becomes fiction.
📦 Tech Inventory
Current stack in a table (AI-Ready, integration) + readiness assessment by dimension.
😣 Pain points
Department pain point matrix — where it hurts and how much it costs today.
📊 Maturity
AI maturity and readiness model on a 1–5 scale.
💡 Assessment feeds the rest
Notice the chain: maturity and pain points provide context for the planning prompts. The prompts aren’t islands—each one builds on what came before ("based on the ... above").
the stack
by department
scale of 1–5
factual
🧭 Planning: what to do and how much it pays
The second group turns the diagnosis into a plan. It's the part the client opens on Monday and executes — phased roadmap, quick wins, and the return calculation.
30/60/90/180/360 Roadmap
Phases with milestones and dependencies. The path from pilot to scale, with dates.
Quick Wins (effort × impact)
Top 10 initiatives under 60 days, under US$50K, and low risk — with week-by-week steps and KPIs.
ROI Calculator
Investment × gains (efficiency, revenue, risk), payback, 3-year NPV, and scenarios (base/conservative/optimistic).
📊 What the ROI Calculator delivers
- •Simple ROI year 1 + payback in months.
- •3-Year NPV with a 10% discount rate.
- •Industry benchmarks (average vs. top quartile) for context.
phased
<60 days
payback + NPV
with numbers
🔩 Implementation: concrete decisions
The third group brings the plan up against operational reality. It's where strategy becomes purchase and usage decisions: what to build, what to buy, where to cut licenses, and which use cases to run in each department.
🏗️ Build vs Buy
Comparison of vendors and frameworks to decide whether to build or buy.
🧾 Consolidate Licenses
Where tools overlap and how much you can save by combining them.
🧩 Use Cases
Library of department-specific AI use cases, ready to prioritize.
✓ When to buy (buy)
- ✓Common problem, already solved by a mature SaaS
- ✓For a small team, short time to value matters
- ✓Isn’t the company’s competitive advantage
✗ When to build (build)
- →It’s the core of the business; it becomes an advantage
- →No vendor meets the specific needs of this use case
- →Sensitive data calls for full control
💡 Consolidated licenses = quick cash win
Consolidating licenses often pays for the entire project: many companies have three tools doing the same thing. Cutting redundancy means immediate savings—and a strong sales argument.
the decision
cut costs
by department
grounded
🛡️ Governance: what keeps the program on track
The fourth group is what separates a one-off pilot from a serious program. Governance is the part that legal and HR demand — and gives the client confidence to scale AI without getting hurt.
AI policy (acceptable use)
What's allowed and what's not: approved tools, permitted data, human oversight.
Data Governance
Privacy, data classification, and compliance (GDPR/LGPD) — the foundation of trust for using AI.
Change Management
Training and adoption manual: technology fails when people don’t come on board.
💡 Governance sells trust
Many people deliver only the “what” (roadmap) and forget the “how to sustain it.” Including policy, data, and change in the package signals maturity—and is what gets the executive committee to approve the investment.
acceptable use
LGPD/GDPR
adoption
serious program
🇧🇷 Resources + adapt to Brazilian Portuguese
The fifth group leaves the client autonomous after delivery: a prompt library for everyday use and a glossary of terms. And here’s the detail that sells in Brazil—adapting the output to Portuguese.
📚 Prompt Library
Starter kit of ready-made prompts by area—the client keeps getting value without you.
📖 Glossary
AI terms explained in business language — gets everyone on the same page with the board.
// adapt to PT-BR: language instruction in Gemini’s context
# INSTRUÇÃO DE IDIOMA (entra antes das Required Sections) Escreva TODO o documento em português do Brasil (PT-BR). Mantenha termos técnicos consagrados em inglês entre parênteses na primeira menção, ex.: "vitórias rápidas (quick wins)". Use R$ e formato de data DD/MM/AAAA. Tom direto e sóbrio.
✓ Adaptation that works
- ✓Language instruction at the top, not at the end
- ✓Technical terms preserved in parentheses
- ✓R$, LGPD, and Brazilian context in the examples
✗ Raw translation
- ✗Translate everything literally (“quick windows”)
- ✗Keep GDPR/USD in a Brazilian package
- ✗Mixing languages in the same paragraph
💡 Edit the prompt > post-translate
Since all prompts share the same structure, you only need to paste the language instruction once in the context_builder and all 15 are produced in Brazilian Portuguese. Adapting them at the source is cleaner than translating the document after it's finished.
client uses it
bring the board up to date
at the source
post-delivery
✅ Module summary
🎯 Mission 4.2 — Adapt 1 prompt and generate the markdown
Choose a deliverable (e.g., quick_wins), adapt it to Brazilian Portuguese, and generate the document for the company in module 4.1.
- Open the original prompt in
synthesis/prompts/and identify Task / Sections / Format. - Add the PT-BR language instruction at the top of the context.
- Ask Claude Code to generate this deliverable using the
research_cache.jsonfrom Mission 4.1. - Check the Markdown: filled-in tables, technical terms in parentheses, R$ instead of US$.
Success: one .md deliverable in Portuguese, structured and based on the research. What you gained: a prompt of your own, in Brazilian Portuguese, ready to enter the Factory pipeline.
Next module:
4.3 — Orchestration, resumption, and diagrams that sell (state.json, resume, and Mermaid → PNG)