Claude Certified Architect – Professional: Independent Exam Prep
By Justin Thompson
Categories: Education, How To
Created and hosted by Justin Thompson. Independent preparation for Claude Certified Architect – Professional: original HarborWorks cases, worksheets and recall practice on service discovery, evidence and team workflows. Not affiliated with or endorsed by Anthropic; personal views, no employer endorsement. No actual exam items or credential guarantee. These episodes cover selected topics, not the full track. Episodes 1–14 follow the July 2026 guide; 15 onward follow the December 7, 2026 guide (70 items, 60 scored, multiple choice, Foundations scope assumed). Check the guide for your exam date.
17 episodes indexed on PodSearch.
Recent episodes
- Calibrate the judge before you move the model
Eight weeks before the drafts' model retires, the grader says the recommended replacement is better everywhere, and the canary week at the service desk says the opposite: reviewers edit thirty-one percent of its drafts instead of eighteen, and the judge now disagrees with the humans on twenty-two…
- Make the retry safe and the result harmless
On Rotterdam's second Monday live, the shift-change rush brings rate-limit errors the wrapper retries, then create-work-order calls that time out after the facilities system has already executed them, and the wrapper retries those too: thirty-seven duplicate work orders by eight o'clock. The same…
- Pick the surface before the architecture
Rotterdam's works council gets a one-sentence promise: requests from Rotterdam staff are processed inside the European Union, and HarborWorks can show it. Six weeks before go-live, three proposals collide: keep the first-party API and "set the region option," move everything onto the hosted agent…
- Leave a decision the next team can change
A new team inherits HarborWorks's assistant from a forty-slide deck and an eleven-thousand-message channel, and within three weeks the drafts' model is deprecated, Rotterdam needs a location answer, the service desk asks whether the ten-second promise still holds, and an engineer proposes removing…
- Make obligations and unequal outcomes reviewable
Rotterdam's works council asks where employee data is processed, how long it's kept and whether it trains models, and a quick answer in the project channel gets all three wrong for the actual configuration; Health and Safety asks whether the assistant is "HIPAA compliant"; and the monthly slice…
- Put the control before the consequence
On a red-team day, a planted vendor reply tells HarborWorks's assistant to close every open Northgate ticket and email the building C asset list to an outside address, and it drafts both: the prompt rule failed, and the pre-send hook never ran because its script had moved. Meanwhile every outgoing…
- Turn traces into an operating decision
Ticket drafts time out at the Monday shift change, the dashboard's five-second average stays green, and the logs record only "ok" and "failed" while storing full prompts, including an HR complaint. Design the per-request record an incident will ask for, with both request identifiers, the effective…
- Diagnose before declaring a winner
Facilities fails its unsupported-claim floor and three fixes are proposed before lunch: a model switch after fifty Friday conversations, a prompt paragraph, and a suspicion about a stale index. Triage every failed case from the trace into instruction, unsupported-evidence, integration and…
- Define evidence before calling it good
Farah reports a ninety-two percent thumbs-up rate and the steering group says ship it, while Finance stopped using the assistant after it cited a superseded expense policy. Define measures across correctness, unsupported claims, speed, spending, safety and security, each with a baseline, a target…
- Connect the service without granting extra authority
Marisol wires the ticket system into the assistant with the token that was handy, and a supervisor who types "As the Facilities administrator" reads an HR complaint filed against him. Choose MCP, a direct API call, a command-line tool or agent communication per connection from the shape of the…
- Make the prompt a governed asset
Beatriz's system prompt grew from twelve lines to thirty-four hundred words, and one Friday edit changed the draft format, collapsed the cache hit rate and made every answer open with "Great question!" Sort the prompt into standing instructions, demonstrations, per-request data and rules that need…
- Buy an acceptable result and prove the service target
Sam's spreadsheet says four hundred requests a day at six seconds each, and Monday's shift change breaks both the ten-second promise and the budget. Define acceptable per request class, price the whole loop with made-up numbers (tokens in and out, round trips, retries, cached prefixes and effort)…
- Is a second agent worth its boundary?
Owen splits the assistant into a coordinator and two specialists, and the first tests return two thousand words, a draft that says "see lookup results," and a "completed" from a coordinator whose specialist never came back. Price each boundary, keep the specialist whose work is verbose, parallel…
- Draw the work before choosing the agent
Priya asks for "an agent that handles whatever people need," and Sam's prototype routes a Shipping printer to Facilities and marks it urgent because the message felt urgent. Draw the path of one request first: input, evidence, processing, output, review and feedback. Then choose a pattern for each…
- Enable the team and measure the workflow
Plan one shared workflow improvement: a clearer pre-review brief for small bug fixes. Define the baseline, ownership, verification and measurement needed to continue, revise or stop, while keeping claims about productivity and operational recovery within the available evidence.Independent…
- Give each question the right evidence
Choose evidence for each claim in a fictional booking question: an approved policy, an applicable manual and current ticket state. Plan retrieval and context that preserve versions, access and freshness, and withhold claims when essential evidence is missing.Independent preparation, not affiliated…
- What must this service improve?
Turn “make support easier” into a decision brief before selecting a model. Define useful outcomes, exclusions, evidence and owners, then decide whether a fictional HarborWorks proposal needs further design, more discovery or a simpler solution.Independent preparation, not affiliated with or…
PodSearch is a free, independent podcast search engine. Read about PodSearch, the privacy policy or contact us.