Claude Certified Developer – Foundations

Not this time. Anthropic's error reference lists 3 causes of a 429: a rate limit, the usage tier's monthly spend cap, and a spend limit on the Claude Code workspace. A spend-cap 429 has no retry-after header and keeps failing until access resumes, so retrying all night only burns time; the job should stop and tell a person. An ordinary rate limit comes with retry-after, and more workers would make it worse. Overload is a different error, 529 overloaded_error. A malformed request returns 400, not 429.

Not this time. The API has 2 rules here: tool results must come in the message right after the assistant's tool call, with nothing between them, and inside that user message the tool_result blocks must come first, with any text after them. Your text in front of the result breaks the second rule. Plain text loses the tool_use_id that ties the result to the call. An extra assistant message breaks the first rule. A 400 means the request itself is invalid, so sending it again fails the same way.

Not this time. Anthropic's models overview says every Claude model ID is a pinned snapshot, including the dateless IDs used from the 4.6 generation on; for older models a short alias pointed at a dated ID, but that no longer applies. So the model behind claude-sonnet-5-5 does not change. What does change is the model you choose: keep the ID in configuration and move to a new one only after your evals pass on it. An hourly eval run guards against something that does not happen. The other 2 options describe IDs that do not work that way on any platform.

Not this time. This is the case Anthropic's prompt caching page walks through. The cache key covers everything up to the breakpoint, and the time in that block makes it different on every request, so each one writes a new entry and none ever finds a match. Put the breakpoint on the last block that is the same across requests, here the end of the policy text. A prompt below the minimum length shows 0 for both fields, not a write every time. Requests seconds apart are well inside 5 minutes. Caching works on ordinary requests.

Not this time. Manual extended thinking with budget_tokens was deprecated on the 4.6 models, and Claude 4.7 and later reject it with exactly this 400. Opus 5.5 uses adaptive thinking, where the model decides how much to think, steered by the effort level. A smaller budget is still the rejected setting. The beta header has no effect on models that reject manual thinking. And newer models do think; they just choose how much themselves. This is the kind of breaking change between model releases the exam asks about: read the migration notes before changing an ID.

Not this time. Anthropic's "Building effective agents" advises the simplest design that works: workflows give predictability and consistency for well-defined tasks, and agents are for when the model has to decide the path. Here the path never changes, so a prompt chain fits, with a check in code after a step (are all the fields there?) before the next runs. Subagents and a coordinator add cost and moving parts the task doesn't need. One large request mixes 4 jobs and gives you no place to check between them. An agent choosing the order would only be choosing the same order every time, at the price of speed and predictability.

Not this time. An instruction in a prompt is followed most of the time, and most of the time is what failed here. A hook is code that runs at a set point whatever the model decides: a PreToolUse hook sees the tool's input before the tool runs, and returning a deny decision blocks the call. Anthropic's SDK documentation uses this very case, blocking writes to .env files, as its example. Louder or repeated instructions are still instructions. A PostToolUse hook runs after the file has already changed, so it can report the damage but not prevent it.

Not this time. Structured outputs guarantee that a response follows the schema in almost every way, but Anthropic's documentation names one gap: the capitalisation of enum and const strings is not guaranteed, and a value can differ from the schema only in case, with no error and no special stop reason. The documented fix is to compare enum values case-insensitively and not to have values that differ only by case. Parsing free text throws away the guarantees you do have. Retrying pays again for a correct answer. Examples may make the slip rarer, but rarer is not never, and the lookup still breaks.

Not this time. Anthropic's tool documentation calls detailed descriptions by far the most important factor in tool performance: what the tool does, when it should and shouldn't be used, what each parameter means and what it returns, in at least 3 or 4 sentences. "Searches orders" doesn't say that it takes a customer or a date range, and "Gets an order" doesn't say it is the one to use when you already have the number. Forcing one tool breaks every request that needs the other. More thinking can't make up for missing information, and the order in which the tools are listed carries none.

Not this time. A key that was public for an hour has to be treated as copied. Anthropic's advice on API keys is to keep them out of code in environment variables or a secrets manager, add .env files to .gitignore, rotate by creating a new key and deactivating the old one, and watch the Console's Logs and Usage pages. Here that means deactivating the key at once, then fixing the leak and checking for use you don't recognise. Rewriting history or making the repository private doesn't take the key back from anyone who already has it. Waiting for the rotation leaves a working key in other hands. Separate keys for each environment would also have kept the damage to production alone.

Not this time. max_tokens as the stop reason means the response hit the limit your code set, so the JSON is cut off, not wrong. That is a problem in the integration, not in the model's answer, and the trace shows it: the failures line up with one stop reason and the longest inputs. Raise the limit (or split long contracts), and check stop_reason before you parse, so a cut-off answer is caught as such. Anthropic's structured outputs page names this case too: a response cut off at max_tokens may not match the schema. A larger model has the same limit, more required fields make the output longer still, and the network isn't involved.