Texte à trous — Anthropic Claude Architect

Before launching the pipeline, the system administrator reviewed the _____ to verify that each specialized subagent had the correct tool restrictions and prompt instructions.

Each time check_inventory returns a low-stock warning, the _____ feeds that result back into the model's context so it can decide whether to trigger reorder_stock or notify_supplier.

Overnight transcription jobs for recorded interviews are offloaded to the _____ to minimize costs, since results are not needed until the following morning.

When the accumulated conversation history, system prompt, and retrieved documents together exceed the model's _____, the accuracy of extractions from the final pages drops significantly.

Once the data enrichment subagent completes its work, the _____ collects its output, formats it, and delegates the next step to the validation subagent rather than letting any subagent self-direct the workflow.

When the file_upload tool fails, the agent reads the _____ field to determine whether to retry immediately, ask the user to fix the input, or escalate to an administrator.

The send_notification tool applies _____ when it receives a 429 rate-limit response, waiting 2 seconds before the first retry, then 4, then 8, but raises an immediate alert on a 401 authorization failure.

After dozens of conversational turns, the model began omitting mandatory compliance disclaimers, a clear sign of _____ that was corrected by re-injecting the key constraints into the active context.

When the payment gateway returns a timeout, the tool response includes _____ set to true so the agent knows the charge was not processed, rather than assuming success from the absence of a protocol exception.

The document_parse tool sets _____ to false when it encounters a corrupted file format, signaling to the agent that retrying the same input will not produce a different outcome.

By implementing the external data connector according to _____, the engineering team ensured that tool-level business errors were clearly separated from protocol-level request failures in every response payload.

The pipeline was redesigned to use _____ for the sentiment analysis and entity extraction tasks, cutting total processing time nearly in half since neither step depends on the other's results.

The security review flagged a risk of _____ because user-submitted feedback strings were being passed directly into agent prompts without any sanitization layer.

The compliance team required that every claim in the generated report include full _____ metadata—document title, page number, and retrieval timestamp—so that auditors could trace each statement back to its source.

The customer support assistant uses _____ to pull the latest product documentation into its context at query time, ensuring its answers reflect the most recent firmware specifications rather than its training data.

After noticing that the same regulatory paragraph was being retrieved and appended on every turn, the team introduced _____ to collapse these repeated passages before they consumed the remaining token budget.

The team decided against a pure _____ approach after realizing it would silently discard the dietary restrictions the user had specified in the first two turns of the conversation.

When the nightly enrichment job failed at record 340 of 900, the team restarted it from the last checkpoint by having the coordinator read the _____ and skip all previously completed records.

To prevent the model from contradicting established lore as the narrative grew longer, the author pinned a _____ at the top of every request containing the immutable character profiles and world-building rules.

Rather than asking the model to infer the user's current filter selections from a long chat history, the application maintains a _____ that is explicitly injected at the start of each new API call.

Due to _____, the risk-scoring subagent had no knowledge of the data-cleaning decisions made by its sibling, so the coordinator had to explicitly summarize those decisions in the risk subagent's invocation prompt.

The architect used _____ to guarantee that validate_schema always executes before transform_data, eliminating a class of bugs caused by the model occasionally reversing the two steps.

Applying the principle of _____, the team replaced a single monolithic customer_action tool with four focused tools—update_address, change_plan, pause_account, and close_account—each accepting only the parameters relevant to its specific operation.

Rather than raising an exception when a requested record is missing, the database lookup function returns a structured _____ with isError set to true and a descriptive message so the agent can decide how to proceed.

When the extracted date field fails the ISO-8601 format check, the pipeline triggers a _____ by appending the specific validation error to the next request so the model can reformat the value correctly.