Start with the code or the application you own
Claude Code gives an engineer a way to explore files, propose edits and run project commands. The Claude API gives an application control over its requests, tools and conversation history. Those starting points lead to different questions. A developer fixing a client needs to understand the failing branch and its test. A team changing an API model also needs to understand which request settings and response-handling assumptions have changed.
Trace one failed request with Claude Code
The Claude Code guide follows a synthetic client that mistakes an HTTP 503 response for ticket data. It asks the assistant to trace what the caller receives, then introduces a small check before the response body is parsed. The result is a patch whose purpose you can explain: successful responses still return a ticket, while an unsuccessful status reaches the caller as an error.
This guide covers terminal setup, access and permissions, with a focused example rather than a claim that the assistant has tested your own project. Anthropic's Claude Code overview describes the broader product. To carry the exercise through two bugs and a final review, open the HTTP errors and ambiguous writes learning path.
A model migration can change the surrounding client
The Sonnet 5.5 report explains why unchanged token rates can accompany a lower vendor-reported task cost. It also identifies concrete integration changes: thinking settings, response-block handling and forced tool choice. Anthropic's current migration guide says Sonnet 5.5 rejects forced any and tool choices. With automatic choice, an answer can arrive without the intended tool call, so the application needs to distinguish an answer from a completed lookup.
Read the Opus 5.5 evaluation article when the question is whether a model preserves an operational exception while working through conflicting instructions. Its illustrative restart case keeps the acceptance rule fixed while comparing configurations. The dated release reports do not establish which model is cheapest or most reliable for your workload. For repeated application work, the prompt-caching comparison adds retries, rejected outputs and review effort to the cost question.
Keep the condition when the conversation gets shorter
The compaction explainer asks what must survive when older exchanges are replaced with a summary. Its conditional restart is a precise example: “after the queue drains” changes when the action is allowed. On-demand compaction lets the application choose when to summarize and continue from the returned block. That gives the application control over the transition, alongside responsibility for preserving later messages and enforcing the actual execution conditions. See the on-demand compaction documentation.
Reconstruct what a browser session did
For Enterprise investigations, the Chrome transcript article explains the contribution of Claude's Compliance API records. A transcript can help locate the submitted action; the destination's records establish whether the change took effect. Chrome-session coverage remains beta and requires the relevant Enterprise Compliance API access. The session-transcript documentation describes the available records.