cert-insights

Study for the Claude Certified Architect - Foundations exam while you work.

Instead of flashcards, this Claude Code plugin turns your real sessions into prep: when a session actually exercises something the exam tests — a hook fires, an MCP tool is called, a subagent spawns, a cache_control breakpoint or tool_choice comes up — Claude asks you one exam-style multiple-choice question about it, at a natural pause. Answers are logged locally per exam domain, and every new session reads that log, so questions (and the depth of the explanations after you answer) steer toward the domains you are weakest in.

What you get

• Impromptu questions grounded in what your session just did, not textbook trivia. At most ~1 per work block; the task always comes first.
• Familiarity tracking: answers append to ~/.claude/cert-insights/answers.jsonl (local only, never leaves your machine). Session start injects your current per-domain score and tells Claude to prioritize weak/unattempted domains.
• /quiz — an on-demand 5-question drill weighted toward your weakest domains (/quiz context drills a single domain; keys below).

What it looks like

You work normally. When your session does something the exam tests, Claude slips in one question at the next natural pause:

[Claude just wired up an MCP server in your project]

Quick one while that sinks in, Tool Design & MCP Integration: Claude calls a tool and the API response has stop_reason: "tool_use". What must your code do next?

A. Retry the request with a higher max_tokens B. Send a new message containing a tool_result block for each tool_use block C. Wait for a webhook with the tool output D. Nothing — the API executes the tool server-side

You: B

Correct — the API never executes tools; your code runs them and returns tool_result blocks referencing each tool_use id. (D is the classic trap.) recorded: tools-mcp correct (1.1)

That recorded: line is the tracking: your answer lands in a local log, and every future session reads it, steering questions toward whatever you keep getting wrong. Want a focused session instead? Run /quiz.

Exam domains tracked

#	Domain	Weight	Record key
1	Agentic Architecture & Orchestration	27%	agentic
2	Tool Design & MCP Integration	18%	tools-mcp
3	Claude Code Configuration & Workflows	20%	claude-code
4	Prompt Engineering & Structured Output	20%	prompting
5	Context Management & Reliability	15%	context

Domains and weights are taken from the official Claude Certified Architect - Foundations Exam Guide, version 1.0, effective July 2026, exam code CCAR-F (reconciled 2026-09-19). The 30 task statements under these domains, the six exam scenarios, and the official out-of-scope list live in plugin/reference/blueprint.md, which the session-start hook references by path rather than inlining, so it is loaded only when a question needs it.

Answers are recorded per domain and per objective (record.sh <domain-key> <correct|wrong> [objective-id]), and both the hook and /quiz prefer objectives with no recorded attempt over ones already covered.

If a newer exam guide disagrees, edit plugin/reference/blueprint.md and the domain list in plugin/hooks-handlers/session-start.sh, then validate with bash plugin/hooks-handlers/session-start.sh | python -m json.tool.

Note: prompt caching implementation detail and token counting specifics are out of scope for this exam, so the plugin no longer asks about them. See CHANGELOG.md.

Install

From GitHub:

claude plugin marketplace add <github-user>/cert-insights
claude plugin install cert-insights@cert-insights

Or from a local clone:

claude plugin marketplace add /path/to/cert-insights
claude plugin install cert-insights@cert-insights

Pause without uninstalling: claude plugin disable cert-insights (do this after the exam — the hook adds ~3.5KB of instructions to every session).

Reset your familiarity log: delete ~/.claude/cert-insights/answers.jsonl.

Windows: the hooks run via bash, so Git Bash must be on PATH (it is with any standard Git for Windows install). macOS/Linux work out of the box.

How it works

Four small parts, no dependencies:

1. A SessionStart hook (session-start.sh) computes per-domain scores and attempted objectives from the answer log and emits quiz-mode instructions as additionalContext — the same mechanism as Anthropic's official explanatory-output-style plugin, which this descends from.
2. A blueprint (blueprint.md) holding the official domains, the 30 task statements and the out-of-scope list. Referenced by path from the hook, so it is read only when a question needs a specific objective.
3. A recorder (record.sh) that Claude runs after each answered question, appending one JSONL line with the domain, the result and the objective id.
4. The /quiz command for deliberate drills.

Disclaimer

Unofficial and unaffiliated with Anthropic. Question quality is Claude's, not a licensed item bank — treat this as ambient reinforcement alongside the official Anthropic Academy course, not as your only prep. Good luck on the exam.