CCA-P start here · Trap gym

CCA-P trap gym

On the exam, two options are usually good practices. One fact in the question decides which fits. Each page below shows a situation with four options; change one fact and watch the winning answer move. Then try three practice questions.

#TrapDomainLines
1Remove it, don't watch it
When an agent holds a tool its role never uses, the answer is to remove the tool, not to log it, alert on it or ask for confirmation before it runs.
D33.1, 5.1
2Code for musts, prompt for shoulds
A rule that must hold every time belongs in code or permissions; a prompt rule makes it likely, not certain. Soft preferences stay in the prompt, and judgement calls go to a person.
D22.2, 5.1, 7.1
3A bigger model is not the fix
When answers are slow, costly or miss things, a bigger model or a bigger context window is rarely the fix; send the right context first, and move up a model only when the task itself needs more reasoning.
D22.1, 2.4, 3.8
4Check the inputs before the model
When answers go wrong, look at what went into the model (the retrieved passages and the data) before blaming or replacing the model.
D44.4, 3.5
5The least autonomy that works
Choose the simplest pattern that solves the problem: one call, then a fixed workflow, then a single agent, and several agents only when the work splits into independent parts worth the cost.
D11.3, 1.4
6Match the search to the question
Vector search matches meaning; exact codes need keyword search, totals need a database query, and wrong-version or wrong-place results need metadata filters.
D33.6, 3.5
7Act with the user's own access
An agent that serves people should reach data with each person's own scoped access, not a shared all-powerful account; use a service identity only when no person is behind the request.
D33.2
8Measure the outcome, not a proxy
A model's self-rated confidence, a thumbs-up rate or an overall average can look healthy while the real outcome is poor; measure the outcome directly, on the slice that matters.
D44.1, 5.3, 5.5
9Evidence before you ship
A few good-looking examples, or a test that changed two things at once, is not evidence: run the eval set first, change one thing at a time, then pick the live test that fits the question.
D44.3, 6.5
10Cache it, don't cut it
When the same long text goes with every request, put it first and cache it; cutting it down or switching to a smaller model gives up quality for a saving that caching gives for free.
D22.5, 4.5
11Agree it in numbers
No LLM system is 100% right, and “fast and accurate” is not a target: agree numbers against today's baseline, with a plan for the errors, and match the design to the value the sponsor named.
D66.3, 1.6
12Speak to the listener
Shape the answer to who is listening and what they must do with it: value, cost, risk and a recommendation for whoever owns the decision; the how for whoever builds and runs it; a written record when the decision will be questioned later.
D66.2, 6.4
13The lightest protocol that fits
Pick the lightest mechanism that meets the reuse, ownership and security needs: a direct API call or CLI for one known call, an MCP server for a capability many AI clients share, and agent-to-agent when another party's agent owns the whole job.
D33.7
14Load it when it's needed
Load content up front only when it is small, stable and needed on almost every request; otherwise keep a short index in context and fetch the detail when a task calls for it.
D33.8, 2.4
15Review where the risk is
Neither review everything nor review nothing: approval before irreversible or high-impact actions, automation plus random audits for low-risk volume, and full review only when every output carries real weight.
D55.3
16Measure fairness, don't prompt it
An instruction telling Claude to be fair proves nothing: measure outcomes across groups on representative data, look for proxies, tell people when AI is involved, and keep a named person accountable.
D55.5