Learning path map
Detailed content
🔎 Live research with Perplexity
Phase 1 of the Factory. Why live research beats static knowledge, how to choose between quick and deep, and how to build the 9 query categories in a ResearchOutput solid — even when a search fails.
Perplexity searches for company facts in real time. Run in June 2026, it pulls information through June 2026—if a better model comes along, the recommendation changes automatically.
It’s what makes the package feel alive and not like an off-the-shelf template. An LLM on its own makes things up; with research, it cites the present.
Live research · facts vs. hallucinations · information freshness.
Quick (~9 queries, sonar model, ~US$0,05) for a fast assessment; Comprehensive (~18 queries, sonar-pro and deep-research, ~US$0,50) for a complete strategy.
It’s the Factory’s cost lever. You choose the depth based on the budget and the client’s value.
Quick mode · comprehensive mode · cost × depth trade-off.
O TemporalContext injects the current year and month into the templates ({current_year}, {current_month_year}) and sets the recency filter (day/week/month/year).
Without a date in the query, the search tool returns old results. Temporal context is what ensures you get “news from this month,” not from 2023.
Time-based placeholders · recency filter · dated query.
Company profile, industry, competition, technology, AI initiatives, regulatory landscape, news, leadership, and investment — each category with a priority and a quick mode flag.
They’re the nine lenses that cover the entire company. Without them, the diagnosis is skewed and incomplete.
QueryCategory · priority · required_for_quick_mode.
If a query fails, the pipeline continues with partial data. Retries with exponential backoff and a 2s pause between calls help avoid rate limits.
A search failure can’t bring down the entire package. Graceful degradation is what keeps the Factory running.
Graceful degradation · retry/backoff · rate limit.
An object with company_name, raw_research, sources e cost. It’s what Synthesis consumes and what becomes research_cache.json.
It’s the contract between Phase 1 and Phase 2. Structured output makes the next stage reliable and cacheable.
ResearchOutput · sources · research_cache.json.
📝 The 15 Deliverable Prompts
Phase 2 of the Factory. The anatomy of a prompt that generates a consulting document and the 15 prompts organized into five groups—Assessment, Planning, Implementation, Governance, and Resources. And how to adapt each one to Brazilian Portuguese.
Every prompt follows the same pattern: Task → Required Sections (with Markdown tables) → Output Format. The context (research + cheat sheets) comes first.
Understanding the template lets you edit any prompt and create new ones without breaking the pipeline.
Task/Sections/Format · Markdown tables · injected context.
Three prompts that diagnose where the company stands: tech_inventory, pain_points e maturity_assessment (scale of 1–5).
It’s the package’s factual foundation. Without an honest assessment, the roadmap is just a guess.
Technical inventory · pain point matrix · AI readiness.
roadmap (30/60/90/180/360), quick_wins (effort × impact, <60 days) and roi_calculator (ROI, payback, 3-year NPV).
It’s the part the client opens on Monday and executes. A plan with numbers is what closes the contract.
Phased roadmap · quick wins · ROI/NPV.
vendor_comparison (build vs. buy), license_consolidation (cut costs) and use_case_library (cases by department).
Turns strategy into concrete purchasing and usage decisions. This is where the plan meets operational reality.
Build vs buy · license consolidation · use cases.
ai_policy (acceptable use), data_governance (data and privacy) and change_management (training and adoption).
It’s what separates a pilot from a serious program. Governance is what legal and HR demand.
AI policy · data governance · change management.
prompt_library (starter prompt kit) and glossary (AI terms). And a language instruction so the output comes out in Portuguese.
Resources make the client self-sufficient after delivery. Adapting the language is the detail that sells in Brazil.
Prompt library · glossary · language instruction.
⚙️ Orchestration, resumption, and diagrams that sell
The glue that connects everything. The orchestrator that confirms each step, the state.json which lets you pick up where you left off when context fills up, render the Mermaid diagrams (current state vs. future state) as PNGs, and package the final download.
Research → Synthesis → Generation, in sequence. Each phase confirms completion before the next begins — nothing moves forward in the dark.
It’s the Factory’s backbone. Confirmation at each stage is what makes the pipeline auditable and resumable.
Pipeline orchestrator · phase confirmation · chained order.
O state.json stores the current phase, the status of each deliverable, and the session's total accumulated cost.
It’s the Factory’s memory. Without a checkpoint, any interruption forces you to redo everything—and pay again.
Checkpoint · state.json · total_cost.
The Claude Code context can fill up (it happened twice during development). You open a new session and run resume "Empresa" — it picks up where it left off.
It’s what lets the Factory survive long sessions. Resuming from the task list is the skill that separates those who finish from those who get stuck.
Context limit · summary · persisted task list.
Two flowcharts: today’s stack (with bottlenecks marked) and the AI-enabled architecture (unified data layer, AI services, automations).
The "before and after" is the slide that sells. Seeing the leap laid out is more convincing than three pages of text.
Current state · future state · marked bottlenecks.
O mermaid_renderer converts the code into a PNG for insertion into the PPTX. Alternatives: the skill beautiful-mermaid and eraser.io.
Client doesn't open the file .mmd; it opens an image on the slide. Rendering is what makes the diagram presentable.
Mermaid → PNG · beautiful-mermaid · eraser.io.
Phase 3 organizes everything into folders (markdown/, presentations/, documents/, mermaid_images/) and wraps up the package for download.
Delivery is half the work. A well-organized package is what makes the client feel they paid for something substantial.
Folder structure · final package · download.