GenAI Professional is scenario-heavy. Treat it like a system-design exam for generative workflows, not a prompt-hack exam.
How to use this plan well
Each study block should do four things:
- review one pipeline layer
- do a short scenario or mixed set
- classify each miss as retrieval, generation, safety, or operations
- route the weak lane into the next block
flowchart LR
Read["Review one layer"] --> Drill["Do short scenario set"]
Drill --> Review["Review why misses happened"]
Review --> Route["Route weak lane into next block"]
Route --> Read
How long should you study?
Typical candidates need 60 to 110 focused hours.
| Your time |
Recommended timeline |
Good fit |
| 16 to 20 hrs/week |
30 days |
intensive path with recent AI or cloud experience |
| 9 to 12 hrs/week |
60 days |
balanced path for most candidates |
| 5 to 7 hrs/week |
90 days |
part-time path with slower reinforcement |
30-day intensive plan
| Week |
Focus |
Output |
| 1 |
LLM basics, prompting, failure modes, capability boundaries |
layer notes and short drills |
| 2 |
RAG: chunking, embeddings, retrieval tuning, metadata filters |
retrieval tie-break sheet |
| 3 |
evaluation, safety, prompt injection, governance |
miss log by failure layer |
| 4 |
deployment, monitoring, cost control, final compression |
mixed sets and final review |
60-day balanced plan
| Phase |
Weeks |
Focus |
| 1 |
1 to 2 |
model basics, inference, prompts, and failure modes |
| 2 |
3 to 4 |
grounding, embeddings, chunking, and retrieval boundaries |
| 3 |
5 to 6 |
evaluation strategy, safety controls, prompt injection defense |
| 4 |
7 |
deployment, serving, latency, and cost trade-offs |
| 5 |
8 |
weak-lane repair |
| 6 |
9 to 10 |
mixed timed review and final compression |
90-day part-time plan
| Month |
Focus |
Goal |
| 1 |
vocabulary, prompting, and model-capability boundaries |
stop losing points to term confusion |
| 2 |
grounding, retrieval, and evaluation |
build stronger system judgment |
| 3 |
safety, deployment, and exam-style tie-breaks |
finish with mixed-set confidence |
If misses cluster here, do this next
| Miss pattern |
Weak lane |
Fix next |
| you keep blaming prompts for retrieval problems |
grounding and retrieval |
review chunking, embeddings, filters, and top-k |
| you trust fluent answers too easily |
evaluation |
review groundedness, correctness, and layer-by-layer scoring |
| you ignore hostile or untrusted input |
safety |
review injection risk, data boundaries, and guardrails |
| you choose a powerful-looking answer that is expensive or hard to operate |
deployment and ops |
review latency, cost, monitoring, and rollback |
What strong prep usually does
- studies by system layer instead of chasing every new GenAI feature
- writes down whether each miss came from retrieval, generation, safety, or operations
- drills confused pairs like grounding vs fine-tuning and model capability vs wrapper
- uses Oracle and OCI docs to settle disagreements, then comes back here for compression
Final 72 hours
| Keep doing |
Stop doing |
| reviewing weak-lane notes and confusion tables |
opening random new model news or tooling |
| rereading the cheat sheet and glossary |
treating prompt tricks as the whole exam |
| running short scenario classification drills |
building a big late lab from scratch |
| checking official OCI docs for disputed boundaries |
trusting unsupported community summaries |
Route yourself well