CCA Foundations diagnostic — 21/25 baseline
First diagnostic taken cold, ~2 weeks without study, on claudecertificationguide.com (independent community resource; 100+ phData engineers have passed Architect Foundations, many citing it as their best mock source — see [[2026-07-30-tool-count-limits-ground-truth]]).
Score: 21/25 (84%). Cert deadline 2026-11-22; this is the starting line, not the finish.
Domain breakdown
| Domain | Score | Band |
|---|---|---|
| D3 Claude Code Configuration & Workflows | 5/5 (100%) | Strong |
| D5 Context Management & Reliability | 5/5 (100%) | Strong |
| D1 Agentic Architecture & Orchestration | 4/5 (80%) | Strong |
| D2 Tool Design & MCP Integration | 4/5 (80%) | Strong |
| D4 Prompt Engineering & Structured Output | 3/5 (60%) | Moderate |
The 100% domains (D3, D5) are the two with daily operational reps — they're what the founder lives in running the COO-agent harness. D4 at 60% is where he directs rather than authors, and it holds 2 of the 4 misses. The tool's own recommended study order also puts D4 first.
The miss pattern — one reflex, not four gaps
Three of the four misses share a single shape:
When a purpose-built mechanism exists at the API or schema layer, he reached past it — for something either softer (prompting) or heavier (restructuring components).
| Q | Domain | His answer | Keyed answer | Direction of error |
|---|---|---|---|---|
| Q01 | D1 · multi-concern handoff | Route to two specialised agents | Decompose within one workflow, investigate in parallel with shared context, synthesise | Added structure |
| Q02 | D2 · tool boundary descriptions | Consolidate two tools into one that routes by intent | Add boundary descriptions to both tools | Removed structure |
| Q03 | D4 · tool_choice forcing | Prompt-based JSON with strict formatting instructions | Force tool_choice: {type: "tool", name: ...} |
Swapped an API guarantee for a prompt tendency |
| Q04 | D4 · few-shot style | 2-4 examples, complete test files + reasoning | 2-4 examples + reasoning covering varied scenarios (async, error handling) | Near miss — coverage nuance only |
Q01 and Q02 look like opposite errors (split vs. merge) but aren't: in both, the keyed answer holds the component topology fixed and fixes the specification — the workflow pattern in one case, the tool descriptions in the other. Q03 is the same instinct one layer down: tool_choice is a guarantee, a prompt instruction is a tendency, and the stem's word "most reliable" was the tell.
This is a data-engineering reflex — reach for the component diagram — that serves him well in pipeline design and misfires in agent design, where the model's natural-language understanding is the routing layer.
Q04 is not part of the pattern. Right technique, right example count; he missed only that examples must span varied scenarios so the model generalises rather than pattern-matching the shown cases.
Where it shows up in live work
The Q01 reflex is one he has already gotten right in production — the Lionsgate AR contract agent is one deterministic flow with the agent contained inside it, not agent-per-concern (founder's own design, 2026-07-29). Right in practice, wrong on the test: this is a test-reading reflex to notice, not a capability gap.
Relevant to CAF too — the Fabric seam work will involve exactly these tool-boundary and decomposition calls.
Recommendation
Study D4 only. The other four domains sit at 80-100% and their misses collapse to one now-visible reflex. Re-test after D4 work to confirm the pattern doesn't recur in the 80% domains.
Related: [[2026-05-27-claude-certified-architect-foundations-study-plan]] · [[2026-06-14-cca-harness-mechanics-cheatsheet]] · [[2026-07-30-tool-count-limits-ground-truth]]