This sheet condenses the principles that the explanations on this site's CCAR-P practice questions keep teaching, one blueprint domain at a time. It is not recalled exam content, which is what exam dumps sell, and it holds no questions. To put the principles to work, take the free CCAR-P practice exam.
Solution Design & Architecture · 17%
Start from the outcome the business measures
- Scenario
- A sponsor names a deliverable, or leaders rank initiatives or judge payoff.
- Principle
- Anchor on the tracked outcome and its baseline; value each initiative by its effect there.
- Tempting trap
- The easiest number to count, like hours saved, often far from the outcome.
- Exception
- Non-language problems get a scoped model slice; unproven gains get a preset-threshold pilot.
Guard the handoff and keep the corrections
- Scenario
- Model output feeds a system of record or later stage; errors surface downstream.
- Principle
- Code checks rules against source data before acceptance; failures go to repair; corrections become labeled cases.
- Tempting trap
- A prompt rewrite or bigger model on a hunch, with no failure record.
- Exception
- A formula or table lookup belongs in code; the model keeps only the messy reading.
Match the pattern to how the path unfolds
- Scenario
- Choosing or reviewing a single call, scripted workflow or tool-using agent.
- Principle
- Least autonomy that works: fixed or auditable steps are a workflow; emergent paths earn an agent.
- Tempting trap
- Autonomy the task never uses, because agents look capable; it adds delay and variation.
- Exception
- High stakes alone do not demand a script; case-by-case trails suit a bounded agent.
Split work along its dependencies
- Scenario
- A job too big or varied for one call is split among workers.
- Principle
- Each unit gets a narrow brief, set output and sole write ownership; only dependent steps wait.
- Tempting trap
- Giving every unit all the source material; contexts fill with noise and facts diverge.
- Exception
- Reliability, not window fit, decides: one call that already handles the input well stays whole.
Claude Models, Prompting & Context Engineering · 13%
Size the model to the constraint that binds
- Scenario
- A workload with known volume, latency, budget and error cost needs a tier.
- Principle
- Start from task difficulty and the binding limit; change tiers only on your own evaluations.
- Tempting trap
- A tier chosen by benchmark or pilot, or a downgrade saving less than errors cost.
- Exception
- Image input or long context needs rule tiers out before price or accuracy.
Match the prompting lever to the failure
- Scenario
- Output drifts in format or scope, rules get argued away, or reasoning fails.
- Principle
- Name the failure first: rules in the system prompt, format in representative examples, reasoning gets room.
- Tempting trap
- A setting that never reaches the failure, like temperature, which changes variety, not obedience.
- Exception
- When code parses output, constrain syntax; field content on odd inputs still needs examples.
Assemble the prompt from parts with different lifetimes
- Scenario
- Prompts swell with retrieved text, tool output, history or copied instructions.
- Principle
- Stable material goes first and unchanged for caching; per-request material is ranked; old history is condensed.
- Tempting trap
- Making room instead of choosing, like a bigger window or another pasted copy.
- Exception
- Keep a fitting document whole when meaning spans it; skip caching prefixes rarely reused.
Integration · 19%
Scope what the agent can reach
- Scenario
- An agent's tool list grows, and so do wrong picks, cost and injection exposure.
- Principle
- Expose only tools workflows use: retire duplicates, gate high-impact actions, fetch rare ones on demand.
- Tempting trap
- Making an oversized catalog cheaper to carry, not smaller; choice and hijack reach stay wide.
- Exception
- A small, steady toolset stays loaded; on-demand discovery would only add a failure point.
Authorize with the caller's verified identity
- Scenario
- A model-driven client acts on a permissioned system for many different users.
- Principle
- Each user's verified identity rides every call; the server applies their rights, never model-written values.
- Tempting trap
- A control beside the permission decision, like a prompt instruction or later logging.
- Exception
- A connection acting for no user gets a narrow service identity, never a person's stand-in.
Let the question decide the retrieval mechanism
- Scenario
- Retrieval returns fragments, near-miss identifiers, wrong-period passages, stale values or wrong totals.
- Principle
- Supply what resemblance cannot with structure: filters before ranking, keyword matching, or system-of-record queries.
- Tempting trap
- Tuning similarity scoring for exactness, completeness or freshness it never delivers.
- Exception
- When the answer is arithmetic over records, compute it in the store's own query.
Choose the thinnest connection that serves the caller
- Scenario
- Reaching a capability by direct call, shared tool server or another team's agent.
- Principle
- Add layers only when used: fixed logic calls directly; many model clients share one server.
- Tempting trap
- A protocol or model in front of deterministic work; extra parts, delay and variable answers.
- Exception
- Delegate to another team's agent when you need its judgment, not its tools.
Set the time budget from how output is used
- Scenario
- An accuracy stage adds delay, or users call a careful pipeline slow.
- Principle
- Weigh error cost against wait cost; measure delay first, often serial steps or held-back output.
- Tempting trap
- Dropping accuracy stages or shrinking retrieval to feel faster when errors cost more.
- Exception
- When nobody waits, run the full pipeline asynchronously or ahead of time.
Watch quality, cost and each dependency separately
- Scenario
- Dashboards look healthy while answers degrade, bills jump or failures go unexplained.
- Principle
- Trace steps per session, score sampled live quality, tag spend by feature, alert per dependency.
- Tempting trap
- Watching only uptime and average latency; a service can be up, fast and wrong.
- Exception
- Logging everything is no fix: strip sensitive fields, sample routine traffic, keep flagged detail.
Evaluation, Testing & Optimization · 16%
A score covers only what the evaluation contains
- Scenario
- A system passes evaluation, then disappoints in production, or one number gates release.
- Principle
- Score every dimension production depends on, mirror its inputs, and check zero-tolerance cases separately.
- Tempting trap
- A blended average as the gate; strength on common cases hides the rare blocking failure.
- Exception
- Curated or synthetic cases rightly add rare classes, if they extend the production mix.
Check the instrument before trusting the result
- Scenario
- A live experiment names a winner, or grader scores decide what ships.
- Principle
- Trust sound method: random, concurrent assignment planned ahead, and graders checked against experts.
- Tempting trap
- Treating a big sample or high grader score as settled; volume cannot fix incomparable groups.
- Exception
- Live tests are not the only gate: held-out offline comparison and shadow runs count too.
Find the failing stage before fixing
- Scenario
- Quality drops after a model change, or answers fail despite the right source.
- Principle
- Fix where failing and healthy traces first diverge; after model changes, retest prompts instead of loosening parsers.
- Tempting trap
- Repairing the familiar part or reverting the latest change on instinct; the fault returns.
- Exception
- Errors rising with input complexity despite solid prompts mark a model limit; route them up.
Cut cost and delay where they concentrate
- Scenario
- Spend or response time must fall while quality on hard steps holds.
- Principle
- Cut the largest cost or delay term without touching hard steps; judge cost per successful outcome.
- Tempting trap
- Cutting what hard steps rely on without evidence, such as a uniform context cap.
- Exception
- When output cannot shrink, stream the most useful part first, if first words arrive fast.
Governance, Safety & Risk Management · 14%
Enforce authority outside the model
- Scenario
- An assistant reads untrusted content or reaches data and actions; rules must survive manipulation.
- Principle
- Put controls beyond the model: identity-scoped data, credentials in the execution layer, checks that fail closed.
- Tempting trap
- Steering the model, like firmer instructions; odds improve, but nothing is guaranteed.
- Exception
- Fixed rule lists miss paraphrased disclosures; add a model-based classifier and act when either fires.
Plan for failures the output will not show
- Scenario
- Pre-launch review: manipulated inputs, invented or stale facts, silent handoff errors.
- Principle
- Wrong output sounds just as sure; rank risks by harm, checking serious ones outside the model.
- Tempting trap
- The output's own signals as proof, like self-reported confidence or agreement after pushback.
- Exception
- If outside content can steer a broadly credentialed part, narrow its reach before launch.
Review counts only if it can catch errors
- Scenario
- People approve AI output, but approvals are near universal, rushed or partial.
- Principle
- Review before irreversible steps; route by consequence, with confidence one input; show sources; make overriding easy.
- Tempting trap
- Adding process without changing reviewer load or view, like written justifications or second approvers.
- Exception
- Scale review back only on measured error rates, by category; log low-impact reversible actions.
Judge compliance and fairness by what the system does
- Scenario
- A design offers an agreement, notice or removed fields as proof of compliance.
- Principle
- Obligations attach to behavior: minimal data per purpose, the governing regime's rules, measured group fairness.
- Tempting trap
- A signed agreement or other document standing in for the actual data flow.
- Exception
- Reuse or automated decisions need your own assessment, not regulator sign-off; fairness needs group data.
Stakeholder Communication & Lifecycle Management · 14%
Test the request before it becomes a promise
- Scenario
- A stakeholder wants a solution, target, demo result or contract term locked in.
- Principle
- Find the problem behind it, make terms measurable, commit only to what tail data supports.
- Tempting trap
- A demo result or average accepted to keep momentum; real use breaks it.
- Exception
- Refusing any number also fails: offer a representative-case range, not an adjusted demo figure.
Recommend, price the alternatives, let the owner decide
- Scenario
- A decision is contested: leaders prefer another path, or requirements conflict.
- Principle
- Take conflicts to their holders first; then recommend, price alternatives, and let the accountable owner decide.
- Tempting trap
- Settling it without showing the trade-off: conceding, splitting the difference alone, or lobbying first.
- Exception
- A recorded decision reopens when new results undercut it; rerun the original criteria openly.
Hand over intent and obligations
- Scenario
- A system passes to a new team, or its sponsor calls it done at launch.
- Principle
- Hand over decision reasons, verifiable outcome criteria and an owned operating loop, proven by a real change.
- Tempting trap
- More diagrams, walkthroughs or author access; receivers still cannot act alone.
- Exception
- A post-launch quality signal helps only with a threshold, a named responder and a review rhythm.
Developer Productivity & Operational Enablement · 7%
Put team conventions and limits in the repo
- Scenario
- An assistant behaves differently per engineer, or loose settings expose destructive commands or secrets.
- Principle
- Version context files and tiered permissions: low-risk actions run, destructive ones wait, secrets stay hidden.
- Tempting trap
- Reminders and training that fade, or lockdowns people route around.
- Exception
- If one person's assistant misbehaves while teammates are fine, check that person's environment first.
Accept AI work only through an independent check
- Scenario
- An agent commits alone, or a diagnosis is trusted because it sounds right.
- Principle
- AI work lands only after an outside check: an owning engineer, workflow gates, telemetry-tested hypotheses.
- Tempting trap
- The generator's own confidence as the gate, or review after changes ran.
- Exception
- Reading, drafting and testing in a contained workspace need no sign-off; gate what reaches shared systems.