Claude API checks pages against source of truth

The project addresses a routine operational gap: published pages are updated by hand while the real data sits elsewhere, so outdated prices or descriptions of removed features can keep facing customers. The developer chose not to rely on a plain text diff after testing it against sample data, where wording variants such as "email support" and "support by email" were flagged as differences and buried the price gaps and missing features that actually needed attention. Extraction on the page side avoids extra dependencies by using html.parser directly rather than BeautifulSoup, and Claude is prompted to return only a JSON array of semantic conflicts.

The developer also frames the problem as a familiar one in data work: comparing a master table with its output destination, such as a BI dashboard or an export. Deciding how much tolerance to allow before something counts as a drift is comparable to setting the acceptable error threshold in an aggregation pipeline, and a threshold that is too strict picks up too much noise while one that is too loose misses genuine gaps.

The outcome hinges on the one step that could not be exercised: because no ANTHROPIC_API_KEY was available in the environment, the live API was never called, and testing relies on a mock that returns fixed drift JSON to confirm the wiring from extractor to detect_drift to report. Whether the tool behaves as intended on real pages is likely to depend on how Claude handles the prompt against actual HTML, so anyone using it in CI should treat that first live run as the real test. The exit-code-1 behaviour and empty-array default suggest the intended home is an automated pipeline rather than a manual review.