CCDV-F flashcards.

The output hit the max_tokens cap you set and was cut off. Raise the limit or continue the response.

The model finished its turn naturally and is not waiting on anything.

Run the requested tool, send its output in a tool_result block with the matching id, and call the model again. Client tools are not run on the server.

The output reached one of the custom stop sequences you set.

model, messages and max_tokens. system, temperature and stream are optional.

In the top-level system parameter. The messages array only takes user and assistant roles.

Set stream to true and render content_block_delta events as they arrive. A complete stream ends with message_stop.

As an image content block in a user message, with a source of type base64 or url.

For large workloads with no latency requirement, at lower cost. Read each request's result separately: a batch is not atomic.

A cache_control breakpoint after a stable prefix, such as a long manual or system prompt, lets later requests read that prefix from cache.

Uploading a document once and referencing it by file id across requests. Referenced files still count as input tokens.

Back off exponentially with jitter and honor the retry-after header. Limits apply per organization, so switching keys does not help.

A malformed request, such as a missing required field or an unsupported parameter. Rate limits are 429, permissions 403 and overload 529.

x-api-key. anthropic-version is a separate required header that pins the API version.

Call the token counting endpoint, which returns the input tokens for the prompt.

Define a tool with a JSON schema for the fields, force it with tool_choice, and validate the arguments that come back.

Prototype on a capable tier, then test whether a smaller tier passes your eval at lower cost.

The smallest tier that passes your eval, because cost and latency dominate at that volume.

A temperature near 0. Caching reuses computation but does not constrain sampling.

The maximum number of output tokens. Input is bounded separately by the model's context window.

Constrain the output's length and format so the model writes fewer tokens.

On hard multi-step problems where accuracy matters more than latency. Thinking tokens are billed as output on every request.

Run the same eval set through both and compare accuracy, cost and latency.

Call the model; on tool_use run the tool, append the result and call again; stop on end_turn or a turn cap.

Append a tool_result block whose tool_use_id matches the tool_use block, which stays in the history.

The agent loop, tool execution, permission checks, hooks and subagents.

A maximum number of iterations, plus a time or cost budget.

Deterministic checks around tool calls, such as blocking a dangerous command.

They run in their own context and return a result to the parent, not their transcript.

One call generates, a second grades the result, and the work is revised until it passes.

As a durable pause that saves progress and resumes when the input arrives.

Before the question, clearly delimited: documents first, question last.

Wrap the document in tags, label it as data, and keep your instructions outside it.

Summarize older turns and keep recent turns verbatim. A summary keeps decisions and state that truncation would drop.

Add examples of input and expected output, including an edge case.

Use an enum in the parameter's JSON schema and list the parameter as required.

The call failed and the content says why, so the model can adapt. The API does not retry client tools.

Requires the model to call some tool this turn without naming which. auto lets it decide; type tool names one.

It says what the tool does, when to use it, what it returns and how it fails.

Sessions and authentication for the remote server. Spawning a subprocess is the stdio transport.

Treat retrieved content as data, keep instructions in the system prompt, and put approval gates on actions that send or delete.

A reviewed list of read-only commands, with approval required for anything that writes.

Rotate the key, remove it from the code, then audit recent usage.

Redact or minimize it, and set a retention policy.

CLAUDE.md, at the repository root or in .claude. settings.json holds permissions and configuration.

claude -p with the prompt as an argument (print mode).

Run a golden test set with automated grading and compare the scores with the old prompt.