Claude Developer Program

Official Syllabus of the Preparatory Course — Sopheria Global Academy (SGA)

Version française

1. Program Overview

Separation of training and certification. The SGA trains; the SGCB certifies. Completing the Claude Developer Program prepares candidates for the CASD™ certification but neither guarantees nor facilitates its award: the certification assessment (certifying project, oral defense, knowledge exam) is administered independently by the SGCB, under identical conditions for all candidates.

Trademark notice. Claude and Claude Code are trademarks of Anthropic, PBC. The SGA and the SGCB are independent entities, not affiliated with or endorsed by Anthropic. The course draws on Anthropic's public documentation.

2. Pedagogy: Why 5 Participants and Individual Coaching

The Claude Developer Program rests on a deliberate pedagogical choice: very small cohorts and intensive individual coaching, rather than a large-scale lecture course.

Cohorts of 5 participants maximum. This cap is not a marketing claim: it is the condition for each 3-hour group session to work as a hands-on workshop rather than a one-way lecture. Every participant regularly presents the progress of their project to the cohort, receives live code review of their own code from the instructor, and learns from the architectural choices — and mistakes — of the four other projects. With five people, no one can remain a spectator.

Two one-hour individual meetings per week. Each participant builds a different project, with their own technical stack, business domain and constraints. The program's eight hours of individual coaching serve precisely that purpose:

• personalization: the instructor applies the CASD™ body of knowledge to the participant's actual project, rather than illustrating with generic examples;
• live code review: annotated reading of the code written between sessions, identification of design flaws, guided refactoring;
• unblocking: resolving hard points during the meeting (API errors, defective agentic loops, invalidated caches) rather than letting the participant lose a week;
• accountability: the cadence of two meetings per week creates an obligation of continuous progress on the project.

The project as the backbone. There are no artificial labs: each week's proving ground is the participant's own project. This project, scoped in week 1 and delivered at the end of week 4, is designed from the outset to meet the specification of Component A (the certifying project) of the CASD™ assessment blueprint — it can therefore be submitted as-is to the SGCB's certification assessment (see section 7).

Explicit preparation for the three assessment components. The course prepares candidates for the three components of the CASD™ assessment — certifying project (60%), oral defense (25%), 30-question knowledge exam (15%) — and week 4 includes a mock exam under real conditions and an individual defense simulation.

3. Mapping of Weeks to the CASD™ Domains

The six domains and their detailed learning objectives are set out in the official CASD™ curriculum (curriculum.md).

4. Detailed 4-Week Program

Week 1 — Foundations and Project Scoping (Domains 1 and 2)

Group session (3 h).

• How an autoregressive LLM works: tokenization, context window, limits (hallucinations, knowledge cutoff); the Claude model family, selection criteria (quality, latency, cost, context), model lifecycle and migration; per-request cost calculation; distribution platforms (Anthropic API, partner clouds) (D1);
• Anatomy of the Messages API: request, response, content blocks, usage, stop_reason; authentication and client configuration; SSE streaming; multi-turn conversations; error handling and retries (D2);
• Live-coding demonstration: application skeleton with streaming, typed error handling and request-ID logging;
• Scoping workshop: each participant presents their project idea; the cohort and the instructor test it against the CASD™ certifying-project specification.

Individual meetings (2 × 1 h).

• Meeting 1 — scoping. Validation of the project scope: use case, value, feasibility within 4 weeks, coverage of the six domains required by the Component A specification; choice of stack (Python or TypeScript) and of the Claude model per route.
• Meeting 2 — first code review. Review of the initialized repository: structure, secrets management, API client with streaming and error handling; unblocking the first calls.

Expected project progress by end of week. Git repository initialized; one-page scoping document (scope, target architecture, mapping to the six domains); working Messages API integration with streaming and error handling; first cost estimate.

Week 2 — Production Prompts and Agentic Loops (Domains 3 and 4 — Part 1)

Group session (3 h).

• Architecture of a production prompt: system prompt, instruction hierarchy, separation of operator channel and user content; XML tagging; few-shot examples; output-format control and structured outputs; prefill and its limits across model generations (D3);
• Reasoning: chain-of-thought, extended/adaptive reasoning, effort tuning, quality/latency/cost trade-offs; prompt iteration: test sets, variant comparison, non-regression; robustness against prompt injection (D3);
• From prompt to agent: decision criteria single call → workflow → agent; agentic loop on the Messages API (stop_reason == "tool_use", tool_result, is_error, iteration bounds and budgets); SDK tool runners; tool-surface design (D4);
• Cohort peer review: each participant submits a key prompt from their project to group critique.

Individual meetings (2 × 1 h).

• Meeting 1 — project prompts. Live rework of the participant's prompts: XML structure, test set, before/after measurement on their own cases.
• Meeting 2 — agentic loop. Code review of the project's agentic loop and tool definitions; stop conditions, tool-error handling, human-approval points.

Expected project progress by end of week. Versioned production prompts with a test set; at least one working tool-equipped agent with a complete loop, safety bounds and tests; prompt iteration log.

Week 3 — MCP, RAG, Caching and Context Management (Domains 4 — Part 2 — and 5)

Group session (3 h).

• Model Context Protocol: host/client/server architecture, primitives (tools, resources, prompts), transports; building an MCP server; server security and authentication; positioning of the Claude Agent SDK and overview of agent-building approaches; multi-agent orchestration, memory and compaction in long-running loops (D4);
• End-to-end RAG pipeline: chunking, embeddings (via specialized providers, as Anthropic does not offer its own embeddings API), vector stores, hybrid search, reranking; RAG vs. long-context trade-offs; native citations for sourced answers; retrieval and groundedness evaluation (D5);
• Prompt caching: prefix invariant, tools → system → messages ordering, cache_control breakpoints, TTL, write/read economics, auditing silent invalidators (D5);
• Workshop: collective diagnosis of a broken cache and of poor RAG retrieval on cases brought by participants.

Individual meetings (2 × 1 h).

• Meeting 1 — project capabilities. Depending on the project: building the participant's MCP server, or designing their RAG/long-context pipeline with citations; live code review.
• Meeting 2 — optimization. Caching audit on the real project (evidence via the usage fields), per-route cost strategy, context management in long conversations or loops.

Expected project progress by end of week. All technical components required by the specification in place: a tool exposed via MCP, a RAG or long-context component with sourced answers, measured prompt caching with cost analysis. The project works end to end; hardening and documentation remain.

Week 4 — Production, Responsible Use and Preparation for the Three Components (Domain 6 and D1–D6 Review)

Group session (3 h).

• Production hardening: secrets management, defense in depth against injection (deterministic validation, allowlists, human confirmation, sandboxing), guardrails and moderation, refusal handling and fallback strategies, resilience (retries, backoff, graceful degradation) (D6);
• Evaluation and observability: eval suites, LLM-as-judge and its limits, non-regression; structured logging, traces, metrics, alerts (D6);
• Responsible use: provider usage policy, personal data (minimization, GDPR), transparency, human oversight, bias, documentation of system limits (D6);
• Mock knowledge exam under real conditions (30 questions, 60 minutes, SGCB format) followed by an annotated correction with per-domain diagnosis;
• Cross-domain review of the six domains: frequent candidate mistakes, guided reading of the SGCB's public scoring rubrics.

Individual meetings (2 × 1 h).

• Meeting 1 — finalization. Final project review against the Component A specification: code repository, README, technical report (10–15 pages), video demonstration (≤ 10 min); responsible-use analysis note; action plan for remaining gaps.
• Meeting 2 — defense simulation. Individual 30-minute mock defense in the SGCB format: live demonstration, technical questions, authorship-verification questions; debriefing and a personalized revision plan for the knowledge exam.

Expected project progress by end of week. Complete, hardened project: automated evals, observability, guardrails, responsible-use analysis; deliverables in the certifying-project format (repository, README, technical report, video) ready to be submitted to the SGCB if the participant registers for the assessment.

5. Weekly Deliverables

Deliverables are reviewed by the instructor during the individual meetings. The SGA issues a certificate of attendance to participants who attended the sessions and submitted the deliverables of all four weeks. This certificate attests to course attendance; it has no effect on the CASD™ certification assessment and confers no title.

6. Technical Prerequisites

• At least 2 years of professional software development experience;
• Proficiency in Python 3.11+ or TypeScript/Node.js 20+ (coaching is provided in both languages);
• Comfort with Git, the command line, HTTP/REST and JSON;
• A personal development machine and your own Anthropic API key (participants bear their own API costs, generally modest at the scale of a training project; the instructor helps budget them from week 1);
• A project idea to build — refined and validated in week 1;
• Recommended: the AIPE™ certification (AI Prompt Engineer) or equivalent prompt-design experience.

A placement interview (30 min) precedes enrollment: it verifies the prerequisites and the viability of the project idea, in order to keep the cohort homogeneous.

7. Relationship Between the SGA Course and the SGCB Certification

The separation between training and certification is strict and structural:

• The SGA trains. It designs and delivers the Claude Developer Program, coaches the participant's project and issues a certificate of attendance.
• The SGCB certifies. It alone administers the three components of the CASD™ assessment — certifying project (60%), oral defense (25%), 30-question knowledge exam (15%) — under the public assessment blueprint (exam-blueprint.md).
• The capstone may become the certifying project. The project built during the course is aligned from week 1 with the Component A specification; a participant who registers for the CASD™ assessment may submit it as-is. It is then evaluated independently by two SGCB assessors against the public rubric — never by the candidate's instructor. No grade, appraisal or recommendation from the course is transmitted to the SGCB or taken into account in the certification decision.
• No certification advantage. Completing the course confers no exemption, bonus or fast track: the certification assessment is identical for all candidates, whether they took the program, another course, or prepared on their own.
• Separate registration. Registration for the CASD™ assessment is made separately, with the SGCB, at the fees and conditions it publishes.

8. Practical Arrangements

© Sopheria Global Institute of Technology — Sopheria Global Academy. "Scientia. Integritas. Futurum." Claude and Claude Code are trademarks of Anthropic, PBC. The SGA and the SGCB are independent organizations, not affiliated with Anthropic.