pack-aws-genai — a quizmill learning pack
Practice for the AWS Certified Generative AI Developer – Professional
(AIP-C01) exam — AWS's first professional-level GenAI credential
(75 questions, 180 min, pass 750/1000, scaled). This pack is a
30-question sample across foundation-model integration, RAG, agents,
prompt management, deployment, observability, and security: three questions
per category, spanning the Easy, Hard and Very hard tiers, and including ten
exam-style scenarios written for the pack.
The full bank — all 386 questions in the same ten categories and three
tiers — ships as a separate pack, aws-genai-full. The app's upgrade card
(on Home, below the practice loop, and in Settings → Packs) links to it, and
the manifest names it (upgrade.packId, with upgrade.unlockUrl as the
service) so a licence key for it unlocks the full bank in place: paste
the key under Packs → Have a licence key?, or open the /unlock#key=…
link you were sent, and the full bank becomes the active pack next to the
sample — no file involved, and progress is kept per pack. A key for any
other pack is refused with the service's own words. The fallback still
works: insert the full bank's bundle file at /packs and tap Make
active. Engines older than the release that added the key entry ignore
the two fields and show the card alone.
The timed exam (75 questions, 180 minutes) is in the full bank only.
Unofficial. Not affiliated with or endorsed by AWS. This is community
study material — treat every question as draft and cross-check against
AWS docs and the official practice set.
Where the questions come from
Adapted (under MIT) from
JeremyEngineer/aws-aip-c01-practice-exam
— 462 community-authored questions with per-option explanations and AWS doc
links. See NOTICE.md for the license and exactly what was
changed. Options were reordered (upstream always lists the correct answer
first), per-option explanations were merged into one teaching explanation
with the source's AWS doc link appended, and 2 items flagged as imprecise
during review were dropped (462 → 460). A further 141 were removed on
2026-10-13 as restatements of other items in the bank (460 → 319): the
upstream Very hard tier repeated the Hard tier almost item for item, and the
easy-off-N items were condensed copies of the hard-bp-N items.
Since 2026-10-21 the per-option verdicts (why each rejected option is
wrong) sit on the option itself, as its explanation, and show under it
once the answer is checked; the question explanation keeps the lead
reasoning and the AWS doc link. On 2026-10-22 the six hard-bp-* items
were removed: upstream's bp-N questions are its copy of the AWS Skill
Builder Official Practice Question Set, which its MIT licence cannot cover
(see NOTICE.md); each slot went to one of the pack's own
scenarios in the same category.
67 exam-style scenarios (source: "original", tag scenario, ids
aws-genai-<tier>-d<domain>-<n>) were written for the full bank in the
official exam's form — a multi-constraint scenario, full-sentence options,
"LEAST operational overhead" / "Select TWO" — and tagged with the official
domain (domain-1 … domain-5). They concentrate on what the community
bank barely covers: testing, validation and troubleshooting, operational
efficiency, integration patterns, and governance. Each one's sourceRef is
the AWS doc page it is grounded in.
The sample here is 30 of those 386, drawn by id. Three of them
(easy-1-4-20, easy-2-1-14, veryhard-1-1-v1) carry corrections made on
2026-10-22 against current AWS documentation (DynamoDB vector search, the
agentcore CLI, Provisioned Throughput's supported-model list).
Verification
An independent adversarial pass checked a stratified sample of 48
questions from the full bank (all task areas, all tiers) against AWS
documentation: 46 confirmed correct, 0 clearly wrong, 2 imprecise (both
removed). The technical claims held up — including fast-moving 2025 services
(Bedrock AgentCore, Strands, application inference profiles, Guardrails).
That is a ~3% sample, so it is a strong signal, not a guarantee; everything
stays reviewStatus: "draft".
On 2026-10-23 all 30 questions in this sample went through a three-pass
automated item review (solve without the key, guess from the options alone,
and a rubric review for the classic item-writing flaws). Every key was
confirmed, but 25 items were rated major for test-wiseness: distractors
nobody would pick and a key noticeably longer than the rest. The options
were rewritten so that each distractor is a real feature misapplied or a
near-miss, every option in an item is about the same length, and the "Read
more" link supports the claim it sits under. On re-review the same judge
rates 0 items fatal or major, 10 minor and 20 clean, still with every key
confirmed. The 30 ids are unchanged, so progress carries over; the same
ids in the full bank still carry the old wording until it is refreshed.
Categories & tiers
Categories follow the source bank's topic taxonomy (10 groups), which is
the most faithful mapping — not the official exam domains. Use the Tier
filter (Easy / Hard / Very hard, from the source's three banks) to drill by
difficulty.
| Category | Sample (Easy / Hard / Very hard) | Full bank (Easy / Hard / Very hard) |
|---|
| Foundation Models: Selection & Inference | 3 (1 / 1 / 1) | 61 (29 / 22 / 10) |
| RAG, Knowledge Bases & Vector Stores | 3 (1 / 2 / 0) | 57 (33 / 22 / 2) |
| Agents (Bedrock Agents, AgentCore, Strands) | 3 (1 / 1 / 1) | 53 (32 / 16 / 5) |
| Data Preparation | 3 (1 / 2 / 0) | 38 (25 / 11 / 2) |
| Security, Guardrails & Compliance | 3 (0 / 2 / 1) | 37 (0 / 30 / 7) |
| Prompt Management & Inference Control | 3 (1 / 1 / 1) | 37 (18 / 16 / 3) |
| Chunking, Embeddings & Retrieval | 3 (1 / 2 / 0) | 35 (22 / 13 / 0) |
| Model Deployment & Developer Tools | 3 (0 / 2 / 1) | 32 (0 / 28 / 4) |
| Well-Architected GenAI Lens | 3 (1 / 1 / 1) | 24 (14 / 7 / 3) |
| Observability & Monitoring | 3 (0 / 2 / 1) | 12 (0 / 11 / 1) |
The Very hard tier is small on purpose: after the restatements were removed
it holds the 15 upstream items that ask something the Hard tier does not,
plus 22 of the exam-style scenarios.
How the sample was chosen
Three questions per category, picked to look like the full bank in
miniature:
- Every tier a category has is shown. Each category with Easy items
contributes one (seven of the ten; Security, Deployment and Observability
have no Easy tier upstream), and every category with a Very hard item
contributes one unless the slot was better spent on a scenario. The sample
is 7 Easy, 16 Hard and 7 Very hard; the bank is Easy 173, Hard 176, Very
hard 37.
- Both kinds of question. 14 are adapted from the community bank and 16
are the pack's own exam-style scenarios — spread over nine of the ten
categories and all five official domains — so the sample shows the
scenario form the full bank is written in. (Until 2026-10-22 the split was
20 / 10, the 20 including six of upstream's
hard-bp items; see below.)
- Both answer formats. Six of the 30 are "Select TWO".
- One concept per question, across the whole sample, not just within a
category: no two of the 30 ask the same scenario with the same key, and
the richest explanations — the ones that say why each rejected option is
wrong and link the AWS doc — were preferred.
Question ids are the full bank's own, so every id in the sample exists in
the full bank. The 14 ids the sample held before 2026-10-13 that were
removed from the bank as restatements (easy-off-1, -3, -4, -6,
-13, -17 and veryhard-1-3-v3, 1-4-v1, 1-5-v10, 1-6-v6, 2-1-v2,
2-2-v4, 2-4-v5, 3-1-v8) were swapped out; progress recorded against
them is simply left behind. On 2026-10-22 six more went the same way, this
time for provenance: hard-bp-4, -9, -11, -12, -16 and -20 are
upstream's copies of the official practice set and are not ours to
redistribute. Their slots went to hard-d1-7, hard-d1-3, hard-d3-9,
hard-d5-10, hard-d2-10 and hard-d5-8 — same category, same tier, each
a scenario written for this pack.
Coverage caveat vs. the official blueprint
The real AIP-C01 exam has five domains: FM Integration/Data/Compliance
(31%), Implementation & Integration (26%), AI Safety/Security/Governance
(20%), Operational Efficiency & Optimization (12%), Testing/Validation/
Troubleshooting (11%). The community bank is heavily weighted to the first
three themes (RAG, FM selection, agents, security) and under-represents
standalone cost/performance optimization and testing/troubleshooting; the
67 exam-style scenarios were written to narrow exactly that gap (25 of them
are Domain 4 and 5), but the full bank is still strong drilling material
rather than a blueprint-balanced mock. For the
authoritative, blueprint-accurate questions, do the AWS Skill Builder
Official Practice Question Set + Official Pretest — the resource real
passers rate as closest to the exam.
Known limitations
- The upstream bank restated itself across tiers — its Very hard items
repeated its Hard items almost verbatim, and
easy-off-N condensed
hard-bp-N. Those 141 restatements were removed from the full bank on
2026-10-13, so 386 is 386 concepts; what that leaves is a Very hard tier of
only 37 items.
- Weak distractors made most items answerable by elimination until
2026-10-23: an automated item review found no wrong keys but flagged 25 of
the 30 for throwaway distractors and keys twice the length of the other
options. The options were rewritten that day (same ids, same keys);
see Verification above.
Use it
npm run pack:use packs/aws-genai # after copying this dir into the engine's packs/
npm run dev # http://localhost:3000
Or point the engine at this directory directly. To deploy as its own app,
the included .github/workflows/deploy.yml builds it onto the quizmill
engine (static app in aws-genai-app/, named after the manifest id) and
publishes to Cloudflare Pages (project aws-genai-cert) once the repo's
CLOUDFLARE_* secrets are set.
