01-projects/phdata

CCDV-F readiness — Claude Certified Developer – Foundations, cold-sit assessment

2026-10-20·assessment·status: active·source: Exam guide PDF shared via iMessage 2026-10-20 (Version 1.0, Effective July 2026, exam code CCDV-F)
phdatacertanthropicccdv-foundationsreadinesscert-prep

CCDV-F readiness: can Ben sit it tomorrow with no prep?

Verdict: coin flip, leaning pass. Roughly 55% to pass tomorrow cold. About 10 hours of targeted prep gets him to about 85%+.

The reason is structural, not a skill gap. The CCA-F exam guide's explicit out-of-scope list (streaming/SSE, vision, prompt caching beyond "it exists", token counting, rate limits/pricing, MCP hosting, auth/key details) reads almost line-for-line like the CCDV-F in-scope list. CCA-F rewarded architecture judgment; CCDV-F adds a raw-API layer that daily Claude Code use never touches. The vault already shows this exact gap: "D3 API specifics" (Message Batches "no clue", tool_choice, prefill) was his recurring CCA-F weak area from the first quiz on 2026-06-10 through practice exam 2 and the 2026-07-30 diagnostic (D4 at 60%, his lowest domain) — full log in [[cert-progress]].

1. Exam facts (quoted from the guide)

Item Guide text
Exam code CCDV-F
Number of items 53
Item format "Multiple-choice and multiple-response items; each item states how many responses to select"
Time limit 120 minutes
Passing score "Scaled score of 720 on a scale of 100–1,000"
Exam fee $125 USD
Validity "12 months from the date the credential is awarded"
Result reporting "Pass/fail with scaled score (100–1,000), plus percent-correct by domain on the score report"
Retakes 14 days after 1st fail, 30 after 2nd, 90 after 3rd; max 4 attempts per rolling 12 months; fee per attempt
Scheduling Cancel/reschedule up to 24 hours before; inside 24h forfeits the fee

Note on "tomorrow": the 24-hour reschedule rule means a booking for tomorrow is effectively locked once made.

Domains and weights (exact)

# Domain Weight ~Items of 53
1 Agents and Workflows 14.7% ~8
2 Applications and Integration 33.1% ~18
3 Claude Code 3.1% ~2
4 Eval, Testing, and Debugging 2.6% ~1
5 Model Selection and Optimization 16.8% ~9
6 Prompt and Context Engineering 11.0% ~6
7 Security and Safety 8.1% ~4
8 Tools and MCPs 10.6% ~6

Sub-skills (exact weights): D1 = Agent Architecture 4.5% · Agent Construction with Claude 5.3% · Agent Patterns and Frameworks 4.9%. D2 = Understanding Requirements 3.4% · Systems Life Cycle 2.8% · Claude API Mechanics 6.8% · Software Engineering Foundations 7.4% · Claude Application Design 8.6% · Configuration Management 4.1%. D3 = Claude Code Operation 3.1%. D4 = Debugging and Error Handling 2.6%. D5 = LLM Fundamentals 5.2% · Technical Fundamentals 6.1% · Model Selection and Tradeoffs 2.7% · Cost and Token Management 2.8%. D6 = Context Engineering 3.8% · Prompt Engineering 4.6% · Output Handling 2.6%. D7 = AI Application Security 3.2% · Guardrails and Safe Deployment 2.3% · Claude Hooks 1.0% · Identity, Secrets, and Key Management 1.6%. D8 = Tool Implementation 4.4% · MCP Server Development 2.1% · Agentic Customization 4.1%.

Shape vs CCA-F: CCA-F was 60 Q, with Claude Code Config at 20% of the exam. Here Claude Code, his strongest ground, is 3.1% (~2 items). The single biggest domain (D2, a third of the exam) is application/API/software engineering.

2. Domain-by-domain: overlap and readiness

CCA-F reference domains: D1 Agentic Architecture 27% · D2 Tool Design & MCP 18% · D3 Claude Code Config 20% · D4 Prompt Eng & Structured Output 20% · D5 Context Mgmt & Reliability 15%.

# Domain (weight) Overlap w/ CCA-F Readiness Why
1 Agents and Workflows (14.7%) High Green (est. 80%) Workflow-vs-agent, supervisor hierarchies, subagents = CCA-F D1 core. Gaps: Agent SDK specifics, "managed agent deployment models (self-hosted vs. Anthropic-hosted)" (Managed Agents: thinking only, never set up), and named frameworks "Strands, LangGraph, PydanticAI" he has not used.
2 Applications and Integration (33.1%) Low-Med Yellow (est. 70%) Requirements + life cycle + config mgmt are SA home turf (green). API Mechanics 6.8% is red: messages, streaming, vision, thinking, caching, Bedrock/Vertex invocation, batch: all CCA-F out-of-scope or his logged weak area. SW Eng Foundations 7.4% (async, refactoring, code review) is yellow for an advisor-mode architect.
3 Claude Code (3.1%) High Green (est. 90%) Daily operator; CLAUDE.md hierarchy, settings.json, skills, headless, auto-mode. Only ~2 items, so worth little.
4 Eval, Testing, and Debugging (2.6%) Med Yellow (est. 65%) Despite the name, it's only "Debugging and Error Handling": error-type identification and recovery choice. Needs API error classes (400/401/403/404/413/429/500/529 overloaded), retry/backoff, integration-layer vs model-output isolation. Note evals proper are not a weighted sub-skill despite appearing in Section 2.
5 Model Selection and Optimization (16.8%) Low Red (est. 60%) Biggest risk. CCA-F explicitly excluded pricing, caching mechanics, and tokenization. Here: sampling/non-determinism, "fast mode, extended thinking, adaptive thinking, effort levels", SDKs + websockets, Opus/Sonnet/Haiku tradeoffs, "breaking behavior changes across model releases", token usage tracking, "prompt caching, cache check-pointing". He uses these via Claude Code but never sets the parameters.
6 Prompt and Context Engineering (11.0%) High Green (est. 85%) CCA-F D4 + D5 combined; context pruning, compaction, subagent isolation are what his harness does. One caveat: Output Handling (structured outputs, defensive parsing) maps to his CCA-F weakest area (tool_choice/prefill misses, D4 diagnostic 60%).
7 Security and Safety (8.1%) Med Green (est. 80%) Prompt injection and untrusted-input handling are a standing operating rule for him; hooks overlap CCA-F (watch the recurring hook-scoping miss: a Write hook does not catch Bash writes); 1Password-wrapper discipline covers secrets. Content policy / IAM / least privilege are SA-native.
8 Tools and MCPs (10.6%) Med-High Yellow (est. 75%) Tool descriptions + built-in vs custom vs Skills vs MCP tradeoffs = CCA-F D2 strength. MCP Server Development 2.1% is red: authoring and deploying a server, resources/tools/prompts primitives, stdio vs remote transports. CCA-F excluded MCP hosting; he installs and uses servers, has not authored one. Client-side vs server-side tools is also new.

Weighted cold estimate: ~73% raw. The 720 scaled cut's raw equivalent is not published; if it sits near 70-72% raw (the CCA-F working assumption), he is on the line. CCA-F practice data (83% cold, then 70%) shows run-to-run variance of 10+ points, so the outcome is genuinely uncertain.

Hands-on API details a heavy Claude Code user would NOT naturally know

Every one of these is in scope on CCDV-F and was out of scope or a logged miss on CCA-F:

Verify exact names and numbers against current docs.claude.com during prep. This list is from working knowledge, not re-checked today.

3. Are the sample questions representative?

The guide does not warn that the samples are easier. It says: "These illustrative items show the style and cognitive level of the exam. They are not drawn from the live item bank."

Skeptical read: they look easier than the real thing. All three samples (batch for overnight cost, untrusted-input isolation for injection, MCP server for reuse) are single-concept items with three obviously wrong distractors ("raise the temperature", "switch to the smallest model regardless of quality"). They test judgment, which is his strength, and none test a parameter name, event order, or error code. Evidence that real exams run harder than their samples: CCA-F practice exam 2 hinged on fine distinctions (layered "most-complete" answers, hook scoping), and on SnowPro GenAI he said "the test was much harder and different than what I studied for" (2026-09-28). Do not calibrate on the samples.

4. Verdict and minimal prep plan

Cold pass probability tomorrow: ~55%. Strong on judgment-style items (D1, D3, D6, D7, most of D8, the design parts of D2 = roughly 55-60% of the exam). Exposed on the parameter-level API surface (D2 API Mechanics + D5 + MCP authoring + debugging = roughly 30% of the exam), exactly where his CCA-F misses clustered.

Prep plan to ~85%+ (about 10 hours), ROI order:

# Topic Hours Method
1 Claude API mechanics: Messages API params, streaming event sequence, vision/PDF input, extended/adaptive thinking config, Bedrock/Vertex invocation, Message Batches 2.5 Write one ~60-line Python script that streams, sends an image, calls a tool, and submits a 3-item batch. Running it teaches more than reading.
2 Prompt caching + cost/token management 1.5 Docs page + add cache_control to the script and read usage before/after; count-tokens endpoint.
3 Model selection + LLM fundamentals: current lineup and tradeoffs, effort, fast mode, adaptive thinking, sampling params, breaking changes across releases, snapshot vs alias pinning 1.0 Models overview + migration notes.
4 Errors, rate limits, stop reasons, retry strategy 0.75 Errors + rate-limits docs; one flashcard sheet.
5 MCP server authoring: primitives, transports, deploy, auth 1.0 Build a 2-tool + 1-resource stdio server and connect it to Claude Code.
6 Agent SDK + Managed Agents + framework one-pagers (Strands, LangGraph, PydanticAI) 1.0 Agent SDK overview; one paragraph per framework on when to use it.
7 tool_choice, parallel tools, client- vs server-side tools, structured outputs 0.5 Refresher; this was his CCA-F repeat miss.
8 Eval + debugging patterns (integration vs model fault isolation) 0.5 Docs "define success / build evals" pages.
9 50-question mixed /cert-study quiz, API-weighted, then re-drill misses 1.0 Same cram-day protocol that worked for CCA-F and SnowPro — quiz built as [[2026-10-20-ccdv-f-quiz-1]], fact sheet as [[2026-10-20-ccdv-f-study-sheet]], ranked links as [[2026-10-20-ccdv-f-study-resources]].

Total: ~9.75 hours. If time-boxed to the minimum, items 1-4 + 9 (6.75h) cover most of the risk (D2 API Mechanics + D5 ≈ 24% of the exam). Keep the CCA-F test-craft rule: on BEST / most-robust stems, choose the layered, most-complete option.

Recommendation: do not book for tomorrow. Book 3-5 days out after the 10-hour pass. A first-attempt fail costs $125 plus a 14-day wait, and the gap is narrow and well-mapped, so a short prep pass is cheap insurance.