Suno AI Music · 2 free songs, no credit card needed arrow_forward Architectural wisdom and exam strategies shared by Claude Certified Architect candidates.
My first practice runs on Domain 1 were brutal because I kept picking 'single agent with many tools' as the safe answer. Once I started asking 'which agent's prompt would actually carry this context?' the orchestration questions clicked.
agentfoxI kept answering based on whether the code was split. The exam cares about whether the context window is split. Once I got that, two or three questions I'd been second-guessing became obvious.
Jordan ParkSeveral Domain 1 scenarios hinge on a single phrase about when the loop should stop. If you skim it, you'll pick the answer that looks architecturally elegant instead of the one that matches the stated termination condition.
mreyesStopped drawing fancy boxes-and-arrows diagrams while studying. Instead I draw the actual sequence of tool calls for each scenario. Made it much easier to spot which step would fail under the exam's stated constraints.
prompt_nomadSpent an evening going through K1.2 articles after I realized I was confusing 'the subagent returns to the parent' with 'the parent loses control'. These are two different failure modes with different mitigations.
kainoeFirst attempt I scored around 60% and Domain 1 was my worst. The fix wasn't more content — it was drilling scenarios until I could predict which failure mode the question was describing from the first two sentences.
Sasha LindbergI kept picking 'make it stateless' as the universal answer. The exam will punish you for that when the scenario requires continuity between turns. Read whether the user is coming back mid-task.
devkai42Before looking at the options, I ask: which component in this scenario is actually holding the plan? That one question removes two wrong answers in almost every Domain 1 item I've practiced.
runtime.rinDomain 2 leans hard on 'what happens when the same tool call fires twice.' If you don't have a quick mental model for idempotency keys vs. natural idempotency, those points are gone.
tinytableI realized I was picking whichever schema had the most flexibility. The exam prefers the schema that constrains the model to the valid set. 'One of five enum values' beats 'any string'.
Priya KrishnanI cloned one of the MCP examples and made it do something trivial — convert currencies. Having touched the transport, schema, and error paths made the entire domain feel concrete.
jb.buildsThe exam has at least one question where the 'right' answer is a single composite tool, and at least one where it's several small tools. The distinguishing factor is whether the model needs to reason between the calls.
schema_samuraiDesigning the happy path is easy. I started every practice question by asking 'how does this tool report a partial failure?' Domain 2 got a lot easier after that.
Mei TakahashiA couple of scenario questions hinge on transport. If the scenario says 'running in a hosted environment across a network', you can cross off the stdio answers without reading further.
r-ochoaI used to view the description as a comment. It's not. The model reads it. Half the 'why is the model not calling this tool' questions are answered by 'the description is vague'.
toolsmith92It's shorter than you think. One pass through the official spec answered three exam questions I'd been guessing on — especially around initialization and capability negotiation.
Lars NordqvistPre-tool-use, post-tool-use, user-prompt-submit, session-start — the names matter on the exam. If you can't tell at a glance which hook fires before a tool call, you'll lose easy points.
hookmasterI kept treating CLAUDE.md as a README for humans. It's instructions the agent reads. Questions about 'why isn't Claude Code following convention X?' almost always trace back to what's or isn't in CLAUDE.md.
Rachel WuMeta, but effective. I pointed it at the knowledge articles in this repo and had it generate questions for me. The domain 3 questions stopped feeling abstract after that.
ebenavidezOne is a reusable prompt template, the other is an isolated context. If you know why you'd pick each one, domain 3 becomes significantly easier.
claude.md.enjoyerI spent a week actually configuring a real Claude Code project instead of just reading. Settings, hooks, subagents, permissions — once you've hit the permission prompt yourself, the scenario questions write themselves.
Ash MorganThe permission mode wording in questions is precise. Don't read 'accept all' and assume it means unrestricted. Read the exact mode name and check the matching behavior.
nia_rtUser settings, project settings, local overrides — I kept forgetting which wins. The exam has at least one question where you have to reason about the merge order.
cli_gremlinTook about 20 minutes. Made multiple exam items obvious because I actually knew what each command did rather than guessing from the name.
Dimitri VolkovA common distractor answer is 'use markdown headers'. For Claude specifically, XML tags are the documented preferred structure. Questions about instruction ambiguity often have an XML-tag answer.
promptzenScored poorly on optimization questions because I kept picking the most detailed answer. Short, structured prompts beat long prose almost every time in the scenarios.
Ayaka TanakaSpent forever thinking caching required identical full prompts. It doesn't — it's prefix-based. Once I internalized that, the 'how would you reduce cost' questions became free points.
xtag_bossRole-play framing is useful but limited. The exam has at least one scenario where the distractor is 'add a role persona'. The real answer is usually examples or structure, not a persona.
lfrancoNot a hard rule, but for scenario questions asking 'how many examples should you include', watch for answers in that range. '10+' is almost always the trap answer.
Zuri OkaforSeveral domain 4 questions hinge on what the next system does with Claude's response. If a script parses it, JSON or XML wins. Don't pick 'natural language' because it sounds friendly.
fewshot_fanRan the same task with two different prompts on my own project. Saw firsthand which choices actually changed the output. Abstract 'which is better' questions turned into 'I've seen this'.
miguel-rThere's a structure in the knowledge articles: context, task, constraints, format, examples. I went through every prompt I had in production and rewrote them that way. Domain 4 felt easy after a week of that.
Olena ShevchenkoAt least one question will give you numbers and ask if they fit in the context window. Don't skip the basic arithmetic — it's free points if you have the window size memorized.
tokenchefTripped on this twice. A question asked about 'max response length' and I picked the full context window size. They're separate limits. Know both.
Kenji HaraSimple sliding window script over a fake conversation. Once I'd had to think about 'which turn do I drop first', the exam's pruning strategy questions were obvious.
t_budgetSome scenarios need jittered backoff, some need immediate retry with a different prompt, some need no retry at all. Read what the failure mode actually is before picking.
nadia.devI'd been guessing on context management questions until I read the pruning strategies article end-to-end. The four strategies are all concrete and easy to remember — I wish I'd started there.
Wren CallahanI kept picking 'cache everything' answers. The exam penalizes that when the scenario has small reused prefixes and big variable suffixes. Know which direction the cost asymmetry goes.
pruner_peteIf a domain 5 question describes a production incident and asks 'what would have helped you diagnose this', the answer is almost always structured logging at tool call boundaries.
gm-tanStarted a Claude Code session and just kept going. Watched the context pressure warnings. Felt the difference between 'close to full' and 'genuinely truncating'. No article explains it like seeing it.
Ingrid SorensenNo posts in this category yet.
"Building sustainable cognitive systems starts with understanding the hidden interfaces between reasoning loops and action execution."