Chapter 03 · CCA-F · all 5 domains
20 anti-patterns to avoid on the CCA-F exam.
Exam distractors are almost always anti-patterns dressed up to look correct. Spot the anti-pattern and you eliminate two or three wrong answers before reading them in detail. Eight are critical (they drop you below 720 if missed), ten are common, two are edge cases.
Exam guide · 4 chapters
OverviewD1 · 27% of CCA-F
Agentic Architectures
5 anti-patterns
AP-LOOP-01Critical
Wrong call
Parsing natural language for loop termination
Pattern-matching on 'I'm done' or 'task complete' to exit an agentic loop. The model's narration is descriptive, not authoritative; it drifts and paraphrases under load.
Right call
Check stop_reason: tool_use vs end_turn
stop_reason is the authoritative signal. tool_use means run the tool and continue; end_turn means the model has finished. Read the field, don't infer from prose.
AP-LOOP-02Critical
Wrong call
Arbitrary iteration caps as the primary stopping rule
max_iterations=10 as the main exit. Legitimate long tasks get cut; broken loops still burn budget to the cap.
Right call
Let stop_reason drive the loop; cap is a safety net
If you hit the cap regularly, the prompt or tool surface is the problem, not the cap.
AP-MULTI-01Common
Wrong call
Subagents inherit the full parent conversation
Forwarding all parent context to every subagent burns tokens, contaminates reasoning and breaks the isolation that makes hub-and-spoke predictable.
Right call
Pass only the task brief; subagent runs clean
Explicit task description plus the narrow inputs it needs. Isolation is the reason to use a subagent.
AP-MULTI-02Common
Wrong call
Sentiment-based escalation routing
Escalating when the user 'sounds frustrated'. Sentiment is not complexity; calm users with policy edge cases get missed.
Right call
Escalate on complexity, policy gaps, confidence signals
Policy gap detected, low retrieval coverage, explicit request for a human, tool failure on retry. Sentiment is secondary at best.
AP-MULTI-03Common
Wrong call
Self-reported confidence as an escalation signal
Asking 'how confident are you?' and routing on the answer. Self-reported confidence is uncalibrated and reflexively high.
Right call
Programmatic checks: schema, coverage, retry counts
Missing required fields, coverage below threshold, ≥2 tool retries, policy hooks fired. External to the model's self-perception.
D2 · 18% of CCA-F
Tool Design + Integration
4 anti-patterns
AP-TOOL-01Critical
Wrong call
18+ tools registered on a single agent
Long tool lists degrade selection. The model confuses similarly named tools and wastes context evaluating irrelevant ones.
Right call
4–5 tools per agent; subagents for more
Narrow surface; orthogonal capabilities go to subagents with their own 4–5 tools.
AP-TOOL-02Critical
Wrong call
Generic error messages ('Operation failed')
{error: 'failed'} gives no signal whether to retry, switch tools, escalate or fall back. The model guesses.
Right call
Structured errors: isError, category, isRetryable
A fixed schema the model can route on: category, retryable flag, retryAfterMs, context.
AP-TOOL-03Common
Wrong call
Silently returning empty results as success
A search that returns [] for both 'no matches' and 'auth expired'. The model proceeds on missing data.
Right call
Distinguish 'access failed' from 'genuinely empty'
Empty-on-success is fine for zero matches. Auth failures and outages must surface as categorised errors.
AP-MCP-01Edge case
Wrong call
Mixing project and user MCP configurations
Same servers in .mcp.json and ~/.claude.json. Resolution order surprises teammates and ships capabilities to CI.
Right call
Project MCP in .mcp.json; user config for personal tooling
Production agents read the committed, reviewable file. User scope never fires in CI.
D3 · 20% of CCA-F
Claude Code Configuration & Workflows
4 anti-patterns
AP-CC-01Critical
Wrong call
Prompt-based enforcement for business rules
'Never refund over $500' in the system prompt is probabilistic. Under load or adversarial input the model violates it.
Right call
PreToolUse hooks for deterministic enforcement
Wrap the refund tool with a hook that errors above $500. Deterministic, auditable, the model never sees the path.
AP-CC-02Common
Wrong call
Ignoring the CLAUDE.md hierarchy
Project rules hardcoded in user scope or ad-hoc prompts. Teammates don't inherit them; CI runs a different rule set.
Right call
CLAUDE.md at repo root; .claude/rules/ for topics
User → project → directory, with @import for topic files. Everything in the repo and reviewable.
AP-CC-03Common
Wrong call
Slash commands with allowed-tools: '*'
All tools by default. A prompt injection in a PR description triggers Bash(rm -rf .).
Right call
Explicit allowed-tools per command
Each command declares a narrow list. Add Bash sparingly and only when it must shell out.
AP-CC-04Common
Wrong call
Same-session self-review of generated code
The session that wrote the code rationalises its own choices instead of flagging them.
Right call
Separate session for review
Generator and reviewer in isolated sessions; the reviewer sees only the final code.
D4 · 20% of CCA-F
Prompt Engineering
4 anti-patterns
AP-PROMPT-01Critical
Wrong call
Vague instructions ('be careful', 'use good judgment')
Soft instructions leave compliance to intuition. On edge cases they resolve to whatever dominated training.
Right call
Explicit criteria with concrete decision rules
'For any amount > $500 or currency ≠ USD, return REQUIRES_REVIEW with {amount, currency, reason}.' Testable and debuggable.
AP-PROMPT-02Common
Wrong call
Few-shot examples omitted on ambiguous tasks
Instruction text alone for classification or extraction. Ambiguity compounds with model variability.
Right call
2–4 few-shot examples
Two typical, one edge case, one negative covers most decision boundaries.
AP-PROMPT-03Critical
Wrong call
JSON output with no schema validation or retry
'Return JSON' and parse the first response. Malformed output silently corrupts data at scale.
Right call
tool_use with a schema plus validation-retry loop
Validate every response; on failure append the specific schema error and retry. Two or three retries catch >99%.
AP-PROMPT-04Edge case
Wrong call
tool_choice = 'auto' when only one tool is correct
The model occasionally narrates in text instead of calling the tool, breaking downstream consumers.
Right call
tool_choice = 'any' or the specific tool
If the pipeline needs the tool result, force the call.
D5 · 15% of CCA-F
Context + Reliability
3 anti-patterns
AP-CTX-01Critical
Wrong call
Progressive summarisation losing case facts
Order numbers, plan tier and prior decisions get summarised into prose and lost; the model regenerates plausible values.
Right call
Pinned case-facts block; summarise only narration
Case facts live in an immutable block at the top of context. Narration compresses freely.
AP-CTX-02Common
Wrong call
Aggregate accuracy masking per-document failures
'94% extraction accuracy' hides a failure on invoices that are 10% of volume and 100% of dollar value.
Right call
Stratified metrics per type, field, confidence band
Report by the dimension that drives business impact.
AP-CTX-03Common
Wrong call
Context degradation in long-running sessions
Hours of conversation without checkpointing. Early details fade, contradictions compound.
Right call
Periodic compaction, scratchpad files, recovery manifests
Write files that survive session reset with a manifest of completed steps so recovery is deterministic.
Ready to test yourself?
Ten questions, two per domain, no sign-up. Most distractors are one of the 20 patterns above.