Free CCDV-F flashcards

A bulk overnight deadline with no user waiting on any single item: what shape of infrastructure requirement does that business requirement produce?

Asynchronous bulk submission with a per-item identifier. The requirement fixes a deadline for the whole set and names no per-item response time, which is the latency-tolerant shape the Message Batches API serves, and the identifier lets each result be matched back to its source item.

I can instruct the model in the system prompt to state the policy revision it used, and sample answers monthly to confirm it complies.

Tempting because the answers would usually carry a revision label, but an instruction is guidance the model can omit under pressure, so the audit trail would depend on generated text rather than on a recorded fact.

Why must an input length guard be measured in tokens rather than in words?

A tokeniser splits text into subword pieces, so ordinary English words often cost about one token each while code, punctuation-dense log lines and non-English words split into several tokens apiece. A word or character ceiling therefore approximates the real budget unevenly and lets dense text past. Counting tokens with a tokeniser measures the same unit the request is bounded by, so the guard and the limit agree.

A character count is finer grained than a word count, so surely it is a stable proxy for the model's input budget.

It is tempting because characters are finer grained than words, but the ratio of characters to tokens still varies by script and by content, so a fixed character ceiling either rejects valid prose or lets dense text through. Only a tokeniser measures the unit the request is bounded by.

Five invoice steps always run in the same order, each step must be individually attributable, and the sequence must be reproducible months later. Where does the control flow belong, and why?

In application code. Agents earn their cost by choosing their own path when the path cannot be known in advance. Here the sequence is known, so coded control flow makes the ordering deterministic and gives every step its own boundary for logging, retry and audit, while Claude is still called at the steps where language understanding is genuinely required.

If an agent can call the same five tools, giving it the steps in its system prompt and letting it decide the order should be equivalent and save me writing orchestration code.

Tempting because a model with the same five tools looks equivalent and needs less orchestration code. It is wrong because the order becomes a model decision that can differ per invoice, which defeats both the per-step failure attribution and the reproducibility requirement.

A run keeps exhausting its context because each tool call returns thousands of lines of markup for ten useful lines. Where is that cured?

This is bloat, and bloat is fixed where the payload enters the conversation. Filter the result in the tool handler so only the useful lines are appended, and discard results whose finding has already been recorded as a note. Both reduce what the request carries without changing the number of calls or the quality of the final report.

Surely I can just let the markup accumulate and rely on compaction to summarise it once the working context nears its limit?

Tempting because compaction does reclaim room in a long conversation, but it is the wrong instrument here: it summarises material that should never have been carried in the first place, and the payloads keep arriving at the same rate.