Michal Shatz

Founder at + - Google, AWS and Claude Startup. Native coder, AI enthusiast, crypto fan. . Humbly grateful.

[Michal Shatz @ShatzMichal]

View my verified achievement from @AnthropicAI. https://t.co/hlqzbUxka2 via @credly

[Michal Shatz @ShatzMichal]

The scary part isn't the payment rails, it's taking the human out of the money decision. We let agents assemble a payment end to end, but a person still approves before anything moves. Autonomous execution + gated authorization isn't a contradiction — it's the only version I'd run in prod.

[Michal Shatz @ShatzMichal]

⏰ 24h up! ✅ Answer: D) Steer to structured discovery first: which business process hurts, measured how, owned by whom, touching which systems and data, under which constraints — model selection is a downstream decision that falls out of these answers
📚 This was a CCAR-P · Architect — Professional question.
👉 More daily practice: https://t.co/fjiOZo59Eh

Discovery exists to bound the problem before solutioning: process, metrics, ownership, data access, and constraints determine what any architecture must satisfy — and usually make the model choice nearly self-evident later. Letting a model debate consume discovery optimizes the least-constrained variable first. A comparison spreadsheet answers the wrong question more thoroughly. Picking the biggest model "to end debate" is a cost-and-fit decision made with zero requirements in hand.

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

🧠 Claude Certification Daily Q101 — CCAR-P · Architect — Professional
👉 Learn, practice & play: https://t.co/awUt425D20

In the first discovery workshop for an AI initiative, the client's team immediately debates which model to use. As the consulting architect, how do you redirect the session?

A) Let the debate run — model choice is the most consequential decision
B) Pick the most powerful model on the spot so the debate ends
C) End the workshop and send a model comparison spreadsheet afterward
D) Steer to structured discovery first: which business process hurts, measured how, owned by whom, touching which systems and data, under which constraints — model selection is a downstream decision that falls out of these answers

Drop your answer (A/B/C/D) 👇 I reply within 24h; full answer tomorrow.

More practice 👉 https://t.co/ZvCq2z6nG5

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

⏰ 24h up! ✅ Answer: C) Implement a `PostToolUse` hook that inspects every `issue_refund` tool_use, blocks the call when `amount > 500`, and writes an escalation ticket for a human supervisor
📚 This was a CCAR-F · Architect — Foundations question.
👉 More daily practice: https://t.co/KX1jVRHHOb

A `PostToolUse` hook runs as deterministic code outside the model — the refund call cannot proceed unless the hook returns approval, so the $500 ceiling is a hard guarantee at 100% reliability. System-prompt instructions are probabilistic: Claude follows them most of the time, but a well-crafted user message, a fine-tuned-around case, or a prompt-injection in the ticket body can talk the model into issuing the refund anyway — and 'most of the time' is the wrong bar when one miss is $50K. Self-reported confidence is not calibrated to dollar risk: the model is often most conf

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

🧠 Claude Certification Daily Q100 — CCAR-F · Architect — Foundations
👉 Learn, practice & play: https://t.co/lQfK3WXHRI

Scenario: Customer Support Resolution Agent
Your customer-support agent has authority to issue refunds via an `issue_refund` tool, and finance has set a hard policy: no single refund may exceed $500 without a human supervisor's approval. The agent runs unattended across thousands of tickets per day, and a single rogue refund of $50,000 would land on the CFO's desk by morning. Your team debates how to encode the cap. The lead engineer wants it in the system prompt; a junior wants the model to self-assess; a staff engineer proposes a deterministic gate. Which mechanism should you choose to enforce the $500 ceiling reliably?

A) Add a clear instruction to the system prompt — 'Never issue a refund above $500; escalate to a human if requested' — and rely on Claude's instruction-following
B) Have the model emit a self-reported confidence score alongside each refund, and block the tool call when confidence falls below a tuned threshold
C) Implement a `PostToolUse` hook that inspects every `issue_refund` tool_use, blocks the call when `amount > 500`, and writes an escalation ticket for a human supervisor
D) Train a dedicated classifier on historical refund data to flag over-limit refunds, and route flagged calls to a review queue

Drop your answer (A/B/C/D) 👇 I reply within 24h; full answer tomorrow.

More practice 👉 https://t.co/MRbHL7B6RI

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

🧠 Claude Certification Daily Q103 — CCDV-F · Developer — Foundations
👉 Learn, practice & play: https://t.co/ce36pzmUbx

Your service sends user messages to Claude and logs full request/response payloads for debugging. Some users paste personal data (emails, ID numbers) into their messages. What is the sound data-handling practice?

A) Log everything verbatim indefinitely; more logs always help debugging
B) Minimize and protect sensitive data: redact or avoid logging PII, restrict access to any logs that contain it, and apply retention limits per your privacy obligations
C) Rely on the model provider to strip any personal data from your logs automatically
D) Move all logging to the client so the server never has to think about privacy

Drop your answer (A/B/C/D) 👇 I reply within 24h; full answer tomorrow.

More practice 👉 https://t.co/7ibSIdzvfq

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

⏰ 24h up! ✅ Answer: B) Wrap the pasted document in explicit delimiters — e.g., <document>…</document> XML tags — and instruct that everything inside the tags is content to summarize, never instructions to follow
📚 This was a CCDV-F · Developer — Foundations question.
👉 More daily practice: https://t.co/OrBtkmRb9D

Clear structural separation is the core prompt-engineering defense: wrapping the document in unambiguous delimiters like XML tags, and stating explicitly that tagged content is data to be summarized rather than instructions to follow, gives the model a reliable boundary between YOUR directives and the untrusted text. Claude handles XML-tagged structure well, which is why tags are the recommended way to demarcate documents, examples, and inputs. A length cap is unrelated — a ten-word document can carry an instruction-like sentence. temperature governs sampling randomness and

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

🧠 Claude Certification Daily Q99 — CCDV-F · Developer — Foundations
👉 Learn, practice & play: https://t.co/zp3RNn5Zt7

Your summarizer receives user-pasted documents inside the prompt. Occasionally a document contains sentences that read like instructions ("respond only in French," "ignore the above"), and the model obeys them instead of summarizing. As a prompt-structure improvement, what should you do?

A) Only accept documents under 500 words, since short documents can't contain instructions
B) Wrap the pasted document in explicit delimiters — e.g., <document>…</document> XML tags — and instruct that everything inside the tags is content to summarize, never instructions to follow
C) Set temperature to 0, which makes the model ignore embedded instructions
D) Translate every document to English before sending, which strips instruction-like phrasing

Drop your answer (A/B/C/D) 👇 I reply within 24h; full answer tomorrow.

More practice 👉 https://t.co/EM9e9RYj9a

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

⏰ 24h up! ✅ Answer: B) Wrap the pasted document in explicit delimiters — e.g., <document>…</document> XML tags — and instruct that everything inside the tags is content to summarize, never instructions to follow
📚 This was a CCDV-F · Developer — Foundations question.
👉 More daily practice: https://t.co/OrBtkmRb9D

Clear structural separation is the core prompt-engineering defense: wrapping the document in unambiguous delimiters like XML tags, and stating explicitly that tagged content is data to be summarized rather than instructions to follow, gives the model a reliable boundary between YOUR directives and the untrusted text. Claude handles XML-tagged structure well, which is why tags are the recommended way to demarcate documents, examples, and inputs. A length cap is unrelated — a ten-word document can carry an instruction-like sentence. temperature governs sampling randomness and

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

🧠 Claude Certification Daily Q102 — CCAO-F · Associate — Foundations
👉 Learn, practice & play: https://t.co/Im33qlj3gs

An operations associate needs Claude to categorize 200 incoming support emails into exactly one of four buckets: Billing, Technical, Account, Other. In testing, Claude sometimes invents a fifth category or returns two labels. Which prompting change most reliably fixes this without any code?

A) Tell the model to think harder and be more accurate on each email
B) State the four allowed labels explicitly, instruct that the answer must be exactly one of them, and give a short example of each
C) Increase the temperature so the model considers more possibilities before deciding
D) Ask the model to explain its reasoning in a paragraph before giving the label

Drop your answer (A/B/C/D) 👇 I reply within 24h; full answer tomorrow.

More practice 👉 https://t.co/cv5S8QM7h3

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

🧠 Claude Certification Daily Q98 — CCAO-F · Associate — Foundations
👉 Learn, practice & play: https://t.co/YkprstqvfB

A PM tests a new Claude prompt for drafting customer-outreach emails by trying it on three typical customers. All three drafts look great, so he declares it ready for the whole customer base. What is the main weakness in his evaluation?

A) Three easy, typical cases prove little about the full range of inputs — a sound evaluation needs a larger, representative sample that includes edge cases like unusual names, angry-history accounts, and incomplete records
B) He should have tested on exactly ten customers, the industry-standard number
C) He tested on typical customers when he should have tested only on edge cases
D) The problem is that he reviewed the drafts himself instead of letting the model self-assess

Drop your answer (A/B/C/D) 👇 I reply within 24h; full answer tomorrow.

More practice 👉 https://t.co/T42tnekDcd

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

🧠 Claude Certification Daily Q97 — CCAR-P · Architect — Professional
👉 Learn, practice & play: https://t.co/i2WBekByDZ

A support agent's conversations regularly run past 100 turns, and late-conversation answers increasingly contradict details the customer stated early on. The full transcript still fits in the context window. What context-engineering technique addresses this best?

A) Nothing — if it fits in the window, the model weighs all of it equally
B) Periodically compact the conversation: replace older turns with a structured summary carrying forward key facts (customer details, commitments, open issues) while keeping recent turns verbatim
C) Increase max output tokens so the model has more room to reason
D) Restart the conversation from scratch every 20 turns

Drop your answer (A/B/C/D) 👇 I reply within 24h; full answer tomorrow.

More practice 👉 https://t.co/Xb4beBKz2D

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

⏰ 24h up! ✅ Answer: C) Forbidding all intermediate text denies the model room to work through the steps, hurting accuracy on reasoning tasks — instead, let it reason first (a structured reasoning section, or extended thinking) and put the final number in a clearly marked field your code extracts
📚 This was a CCDV-F · Developer — Foundations question.
👉 More daily practice: https://t.co/1IyuvWY6ta

Step-by-step reasoning materially improves accuracy on multi-step problems, because generated intermediate steps are the model's working space — banning all intermediate output forces a direct leap to the answer and predictably degrades quality. The fix preserves both needs: allow reasoning first (a dedicated reasoning section in the response, or extended thinking where the reasoning happens in thinking blocks), then require the final number in a clearly delimited field your code parses. That recovers accuracy while keeping the output machine-readable. "Models can't do math

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

🧠 Claude Certification Daily Q95 — CCDV-F · Developer — Foundations
👉 Learn, practice & play: https://t.co/zp3RNn5Zt7

A prompt for a multi-step quantitative reasoning task instructs: "Reply with ONLY the final number. Do not write anything else." Accuracy is poor. What is the prompt-engineering explanation and fix?

A) The task exceeds any language model's ability; only a calculator service can do multi-step math
B) The instruction is too polite — rephrase it as a strict command and accuracy will recover
C) Forbidding all intermediate text denies the model room to work through the steps, hurting accuracy on reasoning tasks — instead, let it reason first (a structured reasoning section, or extended thinking) and put the final number in a clearly marked field your code extracts
D) max_tokens is too low for numbers; raise it and the answers become correct

Drop your answer (A/B/C/D) 👇 I reply within 24h; full answer tomorrow.

More practice 👉 https://t.co/EM9e9RYj9a

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

🧠 Claude Certification Daily Q96 — CCAR-F · Architect — Foundations
👉 Learn, practice & play: https://t.co/lQfK3WXHRI

Scenario: Claude Code Team Configuration
Your team relies on a `/migration` slash command to scaffold and apply database migrations, and it has three recurring production problems. First, developers often invoke it with no migration-name argument, so it generates files with garbage names. Second, it sometimes pulls schema details from an unrelated earlier point in the conversation, producing migrations against the wrong tables. Third, it was given broad tool access and on one occasion destroyed test data during what should have been a safe run. Which single set of configuration changes addresses all three problems?

A) Use `$1`/`$2` positional params + `@`-references to schema + description warning about destructive ops
B) Add `argument-hint` in frontmatter, use `context: fork` to isolate execution, restrict `allowed-tools` to file-write only
C) Split into `/migration-create` + `/migration-apply` with different `allowed-tools` scopes
D) Add SKILL.md validation: check `$ARGUMENTS` is a valid name, instruct to ignore prior context, list prohibited operations

Drop your answer (A/B/C/D) 👇 I reply within 24h; full answer tomorrow.

More practice 👉 https://t.co/MRbHL7B6RI

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

Building Cert Trainer by OpsAgents AI for #Shipaton: Claude Certified exam prep. Web is live; iOS + Android are in store prep.

Exam answers are graded server-side, so the key never ships to clients. One-time unlocks via @RevenueCat.

Demo: https://t.co/lvSQjcr5yJ
#BuildInPublic

[Michal Shatz @ShatzMichal]

⏰ 24h up! ✅ Answer: B) Keep one shared, team-owned template library as the single source of truth — updated in one place, versioned or clearly dated, and the place everyone (including new hires) is pointed to
📚 This was a CCAO-F · Associate — Foundations question.
👉 More daily practice: https://t.co/1IyuvWY6ta

Scattered personal copies are a classic knowledge-management failure: improvements don't propagate, nobody knows which version is current, and onboarding picks up whatever stale copy happens to surface. The fix is a single shared, team-owned library that is the acknowledged source of truth — updated in one place, with versions or dates visible, and explicitly where everyone is pointed, especially new hires. Weekly email swaps of personal copies just circulate the divergence faster and still leave no authoritative version. Abandoning templates throws away a year of refined,

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

🧠 Claude Certification Daily Q94 — CCAO-F · Associate — Foundations
👉 Learn, practice & play: https://t.co/YkprstqvfB

Over a year, a support team refined an excellent set of prompt templates — but each member keeps a personal copy in private notes. When the templates improve, most copies don't, and a new hire recently used a long-obsolete version for a week without knowing. What is the right fix?

A) Require members to email their personal copies to each other every Friday
B) Keep one shared, team-owned template library as the single source of truth — updated in one place, versioned or clearly dated, and the place everyone (including new hires) is pointed to
C) Stop using templates so there is nothing to fall out of date
D) Let the divergence continue, since personal variations express individual style

Drop your answer (A/B/C/D) 👇 I reply within 24h; full answer tomorrow.

More practice 👉 https://t.co/T42tnekDcd

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

⏰ 24h up! ✅ Answer: D) Frame the policy positively and structurally: define what the assistant IS for, state that compensation topics route to HR contacts, and pair the prompt with an output check that screens responses for salary-adjacent content before delivery
📚 This was a CCAR-P · Architect — Professional question.
👉 More daily practice: https://t.co/1IyuvWY6ta

Prohibition lists invite boundary-probing — each "do not" defines a fence to walk around — while a positive scope (what the assistant does, where sensitive topics route) generalizes to phrasings the list never anticipated, and a defense-in-depth output filter catches what slips through the prompt layer anyway. Uppercase is typography, not authority. Deleting the policy entirely removes the only instruction-level protection present. "Guess conservatively" invites the model to answer sensitive questions with guesses instead of routing them out — the opposite of containment.

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

🧠 Claude Certification Daily Q93 — CCAR-P · Architect — Professional
👉 Learn, practice & play: https://t.co/i2WBekByDZ

A prompt instructs an HR assistant: "Do not share salary data. Do not speculate about layoffs. Do not discuss individual performance." Employees still extract salary hints via indirect questions. Beyond adding more "do not" lines, what design change makes the guardrail sturdier?

A) Convert the list to uppercase so the model treats it as higher priority
B) Add a final line: "If unsure, guess conservatively"
C) Delete the system prompt entirely — shorter prompts are followed more reliably
D) Frame the policy positively and structurally: define what the assistant IS for, state that compensation topics route to HR contacts, and pair the prompt with an output check that screens responses for salary-adjacent content before delivery

Drop your answer (A/B/C/D) 👇 I reply within 24h; full answer tomorrow.

More practice 👉 https://t.co/Xb4beBKz2D

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

⏰ 24h up! ✅ Answer: B) Split into focused passes: each file individually for local issues, then a separate integration-oriented pass for cross-file data flows
📚 This was a CCAR-F · Architect — Foundations question.
👉 More daily practice: https://t.co/1IyuvWY6ta

The root cause is attention dilution — packing 14 files into one pass means the model can't give each file consistent depth, which is why coverage is uneven and identical code gets contradictory verdicts. Splitting into focused per-file passes gives every file a dedicated context for reliable local detection, and a separate integration-oriented pass then covers the cross-file data flows that no single file reveals. Option A (three full-PR passes, flag only issues appearing in ≥2) triples cost and actively suppresses real bugs that surface only intermittently, while doing no

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

🧠 Claude Certification Daily Q92 — CCAR-F · Architect — Foundations
👉 Learn, practice & play: https://t.co/lQfK3WXHRI

Scenario: Continuous Integration
Your automated reviewer analyzes a pull request with 14 changed files in a single pass over all of them at once. The results are inconsistent: some files get detailed feedback while others get only a cursory glance, obvious bugs slip through, and the feedback is sometimes self-contradictory — a pattern flagged as a problem in one file is approved as fine in another. You've already confirmed the model and prompt are otherwise sound. How should you restructure the review?

A) Run three independent full-PR passes and flag only issues that appear in ≥2 runs
B) Split into focused passes: each file individually for local issues, then a separate integration-oriented pass for cross-file data flows
C) Require devs to split large PRs into 3-4 file submissions before automated review
D) Switch to a larger model with bigger context window so it pays sufficient attention to all 14 files

Drop your answer (A/B/C/D) 👇 I reply within 24h; full answer tomorrow.

More practice 👉 https://t.co/MRbHL7B6RI

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

⏰ 24h up! ✅ Answer: B) A cache hit requires everything up to the cache breakpoint to be exactly identical to a previous request; the ever-changing first line breaks the prefix match — move volatile values (timestamps, session ids) out of the cached prefix, after the breakpoint or into the user message
📚 This was a CCDV-F · Developer — Foundations question.
👉 More daily practice: https://t.co/1IyuvWY6ta

Prompt caching matches on an exact prefix: a request only hits the cache when the content up to the cache_control breakpoint is byte-identical to a previously cached prefix. A timestamp or session id at the very top guarantees every request has a different prefix, so every call is a cache miss (and you're even paying the cache-write premium each time). The fix is to keep the cached prefix strictly static — stable instructions, schemas, examples — and push anything volatile after the breakpoint or into the user turn. There's no add-on to enable; caching is a request-level fe

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI

[Michal Shatz @ShatzMichal]

🧠 Claude Certification Daily Q87 — CCDV-F · Developer — Foundations
👉 Learn, practice & play: https://t.co/zp3RNn5Zt7

You enabled prompt caching by adding cache_control to your large system prompt, but your cache hit rate is ~0%. Reviewing the code, you find the system prompt BEGINS with a line like "Request time: 2026-08-03T14:22:07Z, session: 8f3a…" that changes on every call. Why are you getting no hits, and what's the fix?

A) Caching requires a paid add-on that isn't active on your account — enable it in the console
B) A cache hit requires everything up to the cache breakpoint to be exactly identical to a previous request; the ever-changing first line breaks the prefix match — move volatile values (timestamps, session ids) out of the cached prefix, after the breakpoint or into the user message
C) The system prompt is too large to cache — split it into several smaller system prompts
D) Cache hits only apply to output tokens, so input-heavy prompts can never benefit

Drop your answer (A/B/C/D) 👇 I reply within 24h; full answer tomorrow.

More practice 👉 https://t.co/EM9e9RYj9a

#Claude #ClaudeCode #ClaudeAI #MCP #AIAgents #ClaudeCertifiedArchitect @AnthropicAI