The anti-patterns are the guide's. The five families are ours. The
CCAR-F guide (v1.0, 2026-07) names
5 anti-patterns across 3
of its 30 task statements. Where the guide names them, they
are the distractor family for that statement — items are written against those exact
shapes rather than invented replacements. Grouping them into the five families below is our
taxonomy, and every quoted line on this page is read out of
blueprint.yml at build time.
Why shapes beat memorising items
CCAR-F presents 4 scenarios drawn from a published bank of 6 — 15 possible draws, 14 of which exercise all 5 domains
across 60 items in 120 minutes. There are
30 published task statements. You will not have drilled every one
of them, and the items are not published anywhere, so some portion of any real sitting is
material you are meeting for the first time.
Item formats are multiple-choice and multiple-response, and our practice items carry at
least four options each. Elimination is therefore worth as much as recall: on an unfamiliar
statement, the wrong options usually still have familiar shapes, because the shapes come from
the same small set of engineering mistakes the guide keeps naming.
The elimination drill
- Read every option before judging any of them. The first plausible one is bait.
-
Name the family out loud for each option. If you cannot name one, that option is a
candidate for the key.
-
Ask what the option guarantees. Enforcement, structure and partition guarantee
things; instructions, caps, summaries and confidence scores do not.
-
Check the survivor against the task statement's key facts. If it contradicts one, you have
mis-eliminated — start again rather than trusting the count.
On a multiple-response item, run the same pass per option: family membership is per option,
not per item.
The five families
Hope over enforcement
hope-over-enforcement The option adds an instruction where a gate belongs. It asks the model to comply instead of making non-compliance impossible.
How the wrong option reads
- the verb is "instruct", "remind", "tell the model to", "add to the system prompt"
- a general quality adjective — be careful, be conservative, be thorough — with no criterion attached
- a policy that must always hold, enforced by wording rather than by code
- a guess where a second identifier would settle the question
What the key does instead
The key installs something deterministic — a prerequisite gate, a hook that intercepts the call, an explicit categorical criterion, or a request for the missing identifier.
Named in the guide
· 1 anti-pattern
Key facts that name the correct move
· 6 cited
-
How multiple customer matches require clarification (requesting additional identifiers) rather than heuristic selection
D5 · 5.2 Design effective escalation and ambiguity resolution patterns
-
The difference between programmatic enforcement (hooks, prerequisite gates) and prompt-based guidance for workflow ordering
D1 · 1.4 Implement multi-step workflows with enforcement and handoff patterns
-
When deterministic compliance is required (e.g., identity verification before financial operations), prompt instructions alone have a non-zero failure rate
D1 · 1.4 Implement multi-step workflows with enforcement and handoff patterns
-
Implementing programmatic prerequisites that block downstream tool calls until prerequisite steps have completed (e.g., blocking process_refund until get_customer has returned a verified customer ID)
D1 · 1.4 Implement multi-step workflows with enforcement and handoff patterns
-
Hook patterns that intercept outgoing tool calls to enforce compliance rules (e.g., blocking refunds above a threshold)
D1 · 1.5 Apply Agent SDK hooks for tool call interception and data normalization
-
The importance of explicit criteria over vague instructions (e.g., "flag comments only when claimed behavior contradicts actual code behavior" vs "check that comments are accurate")
D4 · 4.1 Design prompts with explicit criteria to improve precision and reduce false positives
Detect over prevent
detect-over-prevent The option notices the failure after it happens — a cap, a retry, a validator, an abort — where the design fix removes the failure mode.
How the wrong option reads
- a tunable backstop presented as the mechanism: "stop after N iterations", "retry up to N times"
- validation bolted onto free-form output instead of a schema that cannot produce it
- a retry aimed at information that was never in the source
- the blast radius is the whole run: one failure, everything stops
What the key does instead
The key changes what is possible — the protocol field that actually signals completion, a constrained tool that validates its own input, a schema that cannot emit malformed output, local recovery with structured propagation.
Named in the guide
· 2 anti-pattern
s
Key facts that name the correct move
· 6 cited
-
The agentic loop lifecycle: sending requests to Claude, inspecting stop_reason ("tool_use" vs "end_turn"), executing requested tools, and returning results for the next iteration
D1 · 1.1 Design and implement agentic loops for autonomous task execution
-
Replacing generic tools with constrained alternatives (e.g., replacing fetch_url with load_document that validates document URLs)
D2 · 2.3 Distribute tools appropriately across agents and configure tool choice
-
Tool use (tool_use) with JSON schemas as the most reliable approach for guaranteed schema-compliant structured output, eliminating JSON syntax errors
D4 · 4.3 Enforce structured output using tool use and JSON schemas
-
The limits of retry: retries are ineffective when the required information is simply absent from the source document (vs format or structural errors)
D4 · 4.4 Implement validation, retry, and feedback loops for extraction quality
-
Having subagents implement local recovery for transient failures and only propagate errors they cannot resolve, including what was attempted and partial results
D5 · 5.3 Implement error propagation strategies across multi-agent systems
-
The importance of validating accuracy by document type and field segment before automating high-confidence extractions
D5 · 5.5 Design human review workflows and confidence calibration
Scale over design
scale-over-design The option adds quantity — more tools, more agents, more context, more passes over the same undifferentiated work — where the fix is partition and scope.
How the wrong option reads
- more of the same thing: add tools to the set, add agents, raise a limit, pass the whole file
- every agent gets every tool, "so it has what it needs"
- one pass over everything, where the work has two different shapes in it
- a cheaper or bigger execution mode chosen without checking whether the work tolerates it
What the key does instead
The key partitions: role-scoped tool sets, non-overlapping subagent scope, per-file passes plus a separate cross-file pass, tool output trimmed to the fields that matter.
Named in the guide
· 0 anti-pattern
s
The guide names no anti-pattern for this shape — it publishes it the other way round,
in the key facts below, as the correct move. Expect it as an option shape rather than
as a quotable warning.
Key facts that name the correct move
· 6 cited
-
The principle that giving an agent access to too many tools (e.g., 18 instead of 4-5) degrades tool selection reliability by increasing decision complexity
D2 · 2.3 Distribute tools appropriately across agents and configure tool choice
-
Why agents with tools outside their specialization tend to misuse them (e.g., a synthesis agent attempting web searches)
D2 · 2.3 Distribute tools appropriately across agents and configure tool choice
-
Partitioning research scope across subagents to minimize duplication (e.g., assigning distinct subtopics or source types to each agent)
D1 · 1.2 Orchestrate multi-agent systems with coordinator-subagent patterns
-
Splitting large code reviews into per-file local analysis passes plus a separate cross-file integration pass to avoid attention dilution
D1 · 1.6 Design task decomposition strategies for complex workflows
-
How tool results accumulate in context and consume tokens disproportionately to their relevance (e.g., 40+ fields per order lookup when only 5 are relevant)
D5 · 5.1 Manage conversation context to preserve critical information across long interactions
-
Batch processing is appropriate for non-blocking, latency-tolerant workloads (overnight reports, weekly audits, nightly test generation) and inappropriate for blocking workflows (pre-merge checks)
D4 · 4.5 Design efficient batch processing strategies
Truncate over structure
truncate-over-structure The option compresses — summarise, shorten, pick one, return nothing — where the fix is to give the information a structure that survives compression.
How the wrong option reads
- progressive summarisation applied to numbers, dates, and commitments
- an error collapsed into a status string, or into an empty result that reads as success
- one value chosen when credible sources disagree
- a handoff or synthesis with no claim-to-source mapping left in it
What the key does instead
The key adds a durable structure: a case-facts block outside summarised history, structured error context, claim-to-source mappings with dates, a scratchpad that outlives the context window.
Named in the guide
· 0 anti-pattern
s
The guide names no anti-pattern for this shape — it publishes it the other way round,
in the key facts below, as the correct move. Expect it as an option shape rather than
as a quotable warning.
Key facts that name the correct move
· 9 cited
-
Progressive summarization risks: condensing numerical values, percentages, dates, and customer-stated expectations into vague summaries
D5 · 5.1 Manage conversation context to preserve critical information across long interactions
-
Why generic error statuses ("search unavailable") hide valuable context from the coordinator
D5 · 5.3 Implement error propagation strategies across multi-agent systems
-
How to handle conflicting statistics from credible sources: annotating conflicts with source attribution rather than arbitrarily selecting one value
D5 · 5.6 Preserve information provenance and handle uncertainty in multi-source synthesis
-
Extracting transactional facts (amounts, dates, order numbers, statuses) into a persistent "case facts" block included in each prompt, outside summarized history
D5 · 5.1 Manage conversation context to preserve critical information across long interactions
-
The "lost in the middle" effect: models reliably process information at the beginning and end of long inputs but may omit findings from middle sections
D5 · 5.1 Manage conversation context to preserve critical information across long interactions
-
Structured error context (failure type, attempted query, partial results, alternative approaches) as enabling intelligent coordinator recovery decisions
D5 · 5.3 Implement error propagation strategies across multi-agent systems
-
The importance of structured claim-source mappings that the synthesis agent must preserve and merge when combining findings
D5 · 5.6 Preserve information provenance and handle uncertainty in multi-source synthesis
-
Structured handoff protocols for mid-process escalation that include customer details, root cause analysis, and recommended actions
D1 · 1.4 Implement multi-step workflows with enforcement and handoff patterns
-
The role of scratchpad files for persisting key findings across context boundaries
D5 · 5.4 Manage context effectively in large codebase exploration
Trust the model's confidence
trust-model-confidence The option treats a self-reported or natural-language signal — confidence, sentiment, prose, an aggregate accuracy figure — as a measurement it is not.
How the wrong option reads
- a control decision read out of assistant prose rather than a protocol field
- confidence thresholds standing in for categorical criteria
- sentiment or tone as a proxy for complexity, risk, or the need to escalate
- an aggregate accuracy number used to justify cutting review
What the key does instead
The key uses the structured signal, or calibrates the soft one first: stop_reason, explicit escalation triggers, severity criteria with concrete examples, per-segment validation, confidence scores calibrated against a labelled set before they route anything.
Named in the guide
· 2 anti-pattern
s
Key facts that name the correct move
· 8 cited
-
Writing specific review criteria that define which issues to report (bugs, security) versus skip (minor style, local patterns) rather than relying on confidence-based filtering
D4 · 4.1 Design prompts with explicit criteria to improve precision and reduce false positives
-
Why sentiment-based escalation and self-reported confidence scores are unreliable proxies for actual case complexity
D5 · 5.2 Design effective escalation and ambiguity resolution patterns
-
Appropriate escalation triggers: customer requests for a human, policy exceptions/gaps (not just complex cases), and inability to make meaningful progress
D5 · 5.2 Design effective escalation and ambiguity resolution patterns
-
Adding explicit escalation criteria with few-shot examples to the system prompt demonstrating when to escalate versus resolve autonomously
D5 · 5.2 Design effective escalation and ambiguity resolution patterns
-
The risk that aggregate accuracy metrics (e.g., 97% overall) may mask poor performance on specific document types or fields
D5 · 5.5 Design human review workflows and confidence calibration
-
Field-level confidence scores calibrated using labeled validation sets for routing review attention
D5 · 5.5 Design human review workflows and confidence calibration
-
Stratified random sampling for measuring error rates in high-confidence extractions and detecting novel error patterns
D5 · 5.5 Design human review workflows and confidence calibration
-
Running verification passes where the model self-reports confidence alongside each finding to enable calibrated review routing
D4 · 4.6 Design multi-instance and multi-pass review architectures
What this page is not
It is not a list of exam questions, and no item on this site is a reconstruction of one. Every
practice item is original, written against a published task statement, and carries that
statement's id — which is why each citation above links straight to the statement it came
from.
Family membership is also not a scoring signal. It tells you what a wrong option was trying to
do; it says nothing about how a real sitting will score you, and nothing here converts a
practice percentage into the 720 cut on the
100–1,000 scale.