Français

title: "Mock exam & final project"
subtitle: "Applied AI — Advanced Level · Session 10 (capstone)"
author: "Yann Isola"
theme: "ink #1A2230 / teal #0F7A6C / copper #B4612A / light-teal #E9F6F3 / bg #F4F7F6"

Slide 1 — Title

Mock exam & final project

Applied AI — Advanced Level · Session 10 — last session
Yann Isola · Preparation Claude Certified Architect

Slide 2 — The 2-hour program

  1. 45 min — Mock exam: 30 questions, real conditions
  2. 10 min — Diagnosis: score by domain, preparation radar
  3. 45 min — Defenses: your final project, 4 min + 2 min of questions
  4. 12 min — Retrospective: the 10 sessions, the reflexes that remain
  5. 8 min — Decision: reserve the exam, or dated revision plan

Slide 3 — The real exam, in numbers ⚠

Parameter Value ⚠
Questions 60 · MCQ (Multiple Choice Questionnaire) + scenarios
Duration 90 minutes → 90 sec/question
Threshold 720/1000 ≈ 72%
Error penalty None → answer all
Domains 5, weighted: 27 / 18 / 20 / 20 / 15%

⚠ Volatile figures — check the official Anthropic page before booking.

Slide 4 — The 5 areas and your sessions

Domain Weight ⚠ Sessions
D1 · Agent architecture & orchestration 27% S3, S4
D2 · Tool design & MCP 18% S2, S5
D3 · Claude Code & workflows 20% S6
D4 · Prompt engineering & structured output 20% S1, S7
D5 · Context & reliability 15% S8

Slide 5 — Mock exam instructions

  • 30 questions · 45 minutes — same pace as the day
  • No documentation, no AI
  • Answer all — an empty box is a point offered at random that we refuse
  • Two passes method: sure first, marked later
  • Support: web simulator (“Mock exam” tab) or paper

Slide 6 — The reading grid (reminder S9)

Faced with a certification MCQ, three questions:

  1. Root Cause — what is the real problem described?
  2. Least effort — which answer deals with minimal cost?
  3. Guarantee or instruction? — does the statement require a mechanism (code) or does it accept an instruction (prompt)?

Three out of four distractors fall at this grid.

Slide 7 — 🕐 MAKE EXAM — 45:00

Here we go.

(This slide remains projected during the test, with the stopwatch.)

Slide 8 — Correction: read your score| Score /30 | Verdict |

|---|---|
| 27–30 | Ready — reserve, review domains < 80% |
| 22–26 | At threshold, insufficient margin — 1 targeted week + 2nd blank |
| 17–21 | Resume domain sessions < 60% |
| <17 | Do not reserve — structured recovery, real API practice |

Absolute rule: an area < 60% = red zone, even with a good total.

Slide 9 — The 3 recurring traps

  1. “Temperature 0 suppresses hallucinations” → false: a hallucination at temperature 0 is just reproducible
  2. “The system prompt guarantees that…” → instruction ≠ guarantee; only the code guarantees
  3. “Switch to the most powerful model” → power does not replace cutting, specification or measurement

Slide 10 — Prefill + stop_sequences: know by heart

Prefill {"role":"assistant","content":"<verdict>"} + stop_sequences:["</verdict>"]
The model generates: compliant</verdict> and I clarify…

Application receives: conforme — nothing else.

  • The prefill text is not repeated in the output
  • Shutdown sequence is not included; stop_reason = "stop_sequence"

⚠ Volatile API behavior — check current documentation.

Slide 11 — Place for defenses

The final project

  • 4 min presentation — unforgiving timer
  • 2 min of questions — at least one difficult
  • Architecture diagram mandatory
  • Cross-assessment: 2 peers + trainer, grid /100

Slide 12 — The /100 grid, at a glance

Criterion Points
1 · Business problem & framing 15
2 · Architecture & justification 30
3 · Tools & integrations 15
4 · Reliability, security, human in the loop 20
5 · Costs, evaluation, operation 10
6 · Clarity of the defense 10

Slide 13 — What the jury wants to hear

✅ “I chose a simple pipeline because the path is known in advance”
✅ “Escalation is applied by code, the prompt only explains it”
✅ “It costs ~€X/month to Y requests/day⚠, including cache”
✅ “This case isn't covered — and here's why it's OK”

Elimination red flag

❌ “The model is intelligent, it will manage”

Slide 14 — The deciding questions

  • “Your orchestrator drops at 3 a.m. — what does the user see?”
  • “Why an agent and not a workflow?” What is unpredictable?”
  • “Show me the exact point where a human approves. Why there?”
  • “How do you detect a drop in quality before complaints?”
  • “What can a malicious user do to your system?”

Slide 15 — 🎤 SUPPORTS*(Background slide during passes — timer visible.)*

Slide 16 — Retrospective: the road traveled

10 sessions. From model user to systems architect.

S1 API → S2 Tools → S3 Agent SDK → S4 Multi-agents → S5 MCP
→ S6 Claude Code → S7 Prompt engineering → S8 Reliability
→ S9 Scenarios → S10 Capstone

Slide 17 — S1–S2: the foundations

S1 · Claude API in depth
→ The API is stateless: everything the model “knows” is you who put it into context.

S2 · Advanced tool use
→ The tool description is a prompt. Tool error is reasoning data.

Slide 18 — S3–S4: agentic

S3 Agent SDK
→ Workflow if the path is known; agent if it depends on the findings.

S4 · Multi-agent architecture
→ We cut to isolate contexts, not to look pretty on a diagram.

Slide 19 — S5–S6: connection to reality

S5 · MCP in depth (MCP = Model Context Protocol)
→ MCP standardizes the connection; the quality of the descriptions remains your job.

S6 · Claude Code & CI/CD
→ CLAUDE.md is an instruction; settings.json is a warranty.

Slide 20 — S7–S8: rigor

S7 · Prompt engineering advanced
→ A prompt is a software component: versioned, tested, measured.

S8 · Context, reliability & provenance
→ Context is a managed resource. Confidence is journalized.

Slide 21 — S9–S10: validation

S9 · Certification scenarios
→ Think like the author of the question: root cause, least effort, guaranteed.

S10 · Capstone
→ An architecture that cannot be defended in 4 minutes is not included.

Slide 22 — The 3 meta-lessons

  1. Instruction ≠ guarantee — the prompt requests, the code imposes
  2. The least architectural effort — call < chain < agent < multi-agents: we only go up one floor upon proof of insufficiency
  3. Measure before believing — evals, golden sets, provenance logs: intuition initiates, measurement decides

Slide 23 — Readiness self-assessment

Web page “Preparation” tab:

  • 5 statements per domain × 5 domains
  • Cross-reference with your mock exam score
  • → Preparation radar + revision plan prioritized by weighting

Complete honestly, alone, 5 minutes.

Slide 24 — Your decision, in writing

Before leaving, one line:“I reserve the exam for the week of ___”
or
“I review [domains] for ___ weeks, 2nd blank on ___, then I reserve”

Slide 25 — Certification logistics ⚠

  • Check on the official page: format, price, languages, terms, ironing — everything is volatile ⚠
  • Morning slot if possible; 15 min technical margin (online monitoring)
  • The day before: reflex sheets only + 10 simulator questions for rhythm — no new content

Slide 26 — Continue after the course

  1. Practice > reading — a personal project with real API calls per week
  2. Structured monitoring — Anthropic API changelog · release notes Claude Code · MCP specification: three sources, one monthly review
  3. Being challenged — today's defense is a permanent exercise in the profession

Slide 27 — What you know how to do now

  • Design a multi-agent system and know when not to do it
  • Specify tools that the model chooses well
  • Govern Claude Code for a team: instructions AND guarantees
  • Write versioned, tested, measured prompts
  • Build reliability, controlled cost and traceability from design

Slide 28 — Final remarks

You entered knowing how to use a template.
You come out knowing how to design systems.

Certification will validate what you know.
What you have defended today, no one needs to certify:
it was already the work of an architect.

Go book that exam. 🎓

Notes: Opening tone — nothing new today, everything checks out. Two tests: alone facing the questions, standing facing the group.

Notes: Announce strict entry timing. The session does not forgive excesses.

Notes: Last repetition of the golden rule: the figures marked ⚠ are always verified at the source.

Notes: D1 weighs more than a quarter of the exam. A collapsed domain cannot be made up for by others.

Notes: Time announcements at 20, 10 and 2 minutes. No substantive questions during the test.

Notes: Project it 30 seconds before starting the timer. This is the last methodological reminder before the test.

Notes: Silence. Announcements at T-20, T-10, T-2. During this time: prepare the order of defenses and distribute the grids.

Notes: Blank = 30 questions, actual = 60 ⚠ → high variance: reason in “score ± 2”.

Notes: Correct the 3 most failed questions in the group here (show of hands). Likely candidates: prefill/stop_sequences, cache invalidation, workflow vs agent.

Notes: Question that almost certainly falls in one form or another. Redo it on the board if necessary.

Notes: Remember that holding for 4 minutes IS an assessed skill (criterion 6). Order of passage on the board.

Notes: Criteria 2 and 4 make up half the mark — like in the real life of an architect.

Notes: Assuming a limit is worth points; hiding it costs money. This is a bias to be corrected before the real arch committees.

Notes: Trainer's bank — pose at least one per defense if the group remains tender.

Notes: 6 min per person, all inclusive. > 7 participants → 3+2 format announced in advance.

Notes: Open the “Course” tab of the web page — interactive map of sessions and their connections.

Notes: Get people to react: which session changed their practice the most? Often S2 — the discovery that descriptions drive everything.

Notes: The duo which weighs 27% of the exam (D1). The workflow/agent rule is THE recurring design question.

Notes: The S6 sentence is probably the most profitable of the entire course in terms of exam points.

Notes: Evals and golden sets (S7) + compaction, cache, traceability (S8): the transition from demo to production.

Notes: Complete: today's defense was the most realistic examination of the course - that of the profession.

Notes: Have them formulated BY the group first (2 min), then project. They must come from them.

Notes: Emphasize: a weak domain at 27% should be revised before a weak domain at 15%. The page does this sorting automatically.

Notes: Identify commitments — this is the only evaluation of the block and the basis of post-training monitoring.

Notes: Monitoring anti-cramming is counter-intuitive but documented: we consolidate, we no longer ingest.

Notes: Everything marked ⚠ in their notes comes from these three sources. The monthly review is the price of the validity of their knowledge.

Notes: Reread slowly. This is the finish photo — mentally compare it to the start photo from session 1.

Notes: End. Remain available 10 minutes later for asides (especially scores < 15/30 — see appendix C of the guide).