The 15 confusions that cost people points
Not knowing a feature is one way to lose a point. Mixing up two features you do know is another, and it is harder to notice, because both options look familiar and the scenario fits only one of them. The pairs below come from writing and reviewing the question banks here. They are the distinctions that the explanations have to spell out again and again. Each one comes with the test that separates the two and a short example. The exam codes show which exams lean on it.
The exam itself
Scaled score or percentage correct. The scaled score, from 100 to 1,000 with a pass at 720, decides pass or fail. The per-domain percentages on the score report are feedback. Scaling adjusts for forms of slightly different difficulty, so 720 is not 72% correct. All five exams.
Certificate of completion or certification. A Claude Academy certificate shows that you finished a public course and passed its quiz. A certification is a proctored exam at Pearson VUE, scored from 100 to 1,000, open only to people at partner companies, and valid for 12 months. All five exams, at registration.
Working in Claude
Incognito chat or de-identified data. Incognito controls what is kept afterwards: no chat history and no memory. It does nothing about what you disclosed in the first place. When a policy says only de-identified data may be analyzed, strip the names and identifiers before you upload, and screen the free-text fields too. CCAO-F.
Uploaded file or connector. An upload is a snapshot, and it goes stale the moment the source changes unless someone replaces it. A connector reads the current source, with your access rights. A Project that keeps quoting last quarter’s price list points to a stale upload. CCAO-F.
Project instructions, knowledge or a Skill. Instructions are standing rules of behavior: tone, format, what to refuse. Knowledge is reference material: the facts to draw on. A Skill is a reusable procedure or document format that loads when the task calls for it. “Always answer in British English” is an instruction, the product catalogue is knowledge, and the monthly report template is a Skill. CCAO-F.
Building with Claude
Workflow or agent. In a workflow, your code decides the next step and Claude performs each one. In an agent, Claude decides the next step. If an auditor needs proof that every check ran in order, that’s a workflow. An agent fits the open-ended exceptions branch, where the steps depend on what turns up. CCDV-F, CCAR-F, CCAR-P, CCDV-P.
Streaming or batch processing. Streaming shows text as it is generated. The user sees something sooner, at the normal price. A message batch costs about half as much and returns results asynchronously, within 24 hours, matched back to requests by a custom ID. Overnight classification belongs in a batch. A check that blocks a developer’s merge doesn’t. CCDV-F, CCAR-F.
Time to first token or total latency. The time to the first token depends mostly on how much input Claude has to process before it starts, and on any thinking. Total latency depends mostly on how much output it writes and on the model tier. Streaming improves the first, not the second. When a waiting service needs the complete answer, shorten the output. CCDV-P.
Prompt caching or compaction. Caching makes a repeated, identical prompt prefix cheaper and faster to process. The context stays exactly the same. Compaction replaces older turns with a summary. That frees space but loses detail, so facts that must survive belong in state your code keeps and re-injects. CCDV-F, CCDV-P, CCAR-P.
CLAUDE.md, settings or a hook. CLAUDE.md is guidance that Claude reads and usually follows. Settings are configuration that Claude Code enforces, permissions included. A hook is your own code that runs on an event every time, whatever the model decides. “Use pnpm” goes in CLAUDE.md, “never edit the migrations folder” is a deny rule in settings, and “run the formatter after every edit” is a hook. CCDV-F, CCAR-F.
Custom tool, Skill or MCP server. A custom tool is an action your application runs, often on behalf of the signed-in user. A Skill is know-how and files that load on demand. An MCP server is a capability that one team builds once and many clients connect to. A refund action tied to the user’s session is a custom tool, the triage procedure is a Skill, and the shared ticketing integration is an MCP server. CCDV-F, CCAR-F.
Access failure or empty result. “The lookup failed” and “the lookup worked and found nothing” need different handling, so a tool has to report them differently. Retrying an empty result loops forever on a real no-match. Treating a failure as empty hides an outage. CCAR-F.
Running Claude in production
Zero data retention or your own logs. Zero data retention governs what the model provider keeps. It says nothing about your own traces, prompts in your logging pipeline, or the analytics tool you send them to. Those need their own redaction and access controls. It is also not data residency, which is about where processing happens. CCDV-P, CCAR-P.
Detection or prevention. Logging, alerts and dashboards tell you that something bad happened. Gates, hooks, allowlists, sandboxes and scoped credentials stop it from happening. When the scenario describes harm that can’t be undone, such as data leaving the building, more logging is the trap. CCDV-P, CCAR-P.
Contaminated or saturated evaluation. A contaminated suite has cases that leaked into tuning or prompts, so its scores are inflated. A saturated suite passes everything, so it can’t show a regression. Both look healthy on a dashboard. The fix is a held-out set refreshed from real production failures, not a stricter threshold on the same cases. CCDV-P, CCDV-F.
How to use this list
Reading the pairs won’t fix them. Practising on items where one member of a pair is the trap will. Each explanation in the free forms names why each wrong option is wrong. When it names one of these pairs, add a line to your own list of what tripped you up.