Anthropic Claude on Google Cloud

Clear prompts reduce the work everything else must do. You write direct, specific instructions, use XML structure and add examples, then tighten them against real inputs.

• Direct and specific instructions
• XML structure
• Examples that teach
• Handling edge-case inputs
• Prompt templates in code
• Comparing revisions

You build A prompt tightened over three revisions against a set of tricky inputs.

You design evaluations systematically: define what good looks like, generate test datasets, run them on Google Cloud, and grade the results with models and with code.

• Defining success criteria
• Test dataset design
• Running evaluations
• Model-based grading
• Code-based grading
• Interpreting scores and deciding what to change

You build An evaluation suite with a graded report for one application prompt.

Claude calls functions you describe. You write tool schemas, handle results across several turns, and use batch, text edit and web search tools in the same conversation.

• Writing tool schemas
• Handling tool results across turns
• Batch tool calls
• The text edit tool
• The web search tool
• Failure handling in tool use

You build An assistant that calls external functions and reports on each call.

Build retrieval in layers. You chunk and embed text, add BM25 search, combine indexes, and improve results with reranking and contextual retrieval.

• Chunking and embedding text
• BM25 lexical search
• Combining multiple indexes
• Reranking
• Contextual retrieval
• Judging retrieval quality

You build A multi-index retrieval pipeline over a document set, with reranking.

Use features that change what an application can do. You apply extended thinking, handle image input, ask for citations and use prompt caching to cut repeated cost.

• Extended thinking
• Image input
• Citations
• Prompt caching
• Combining features in one request
• Cost and latency effects

You build A multimodal question-answering tool that uses thinking, images and citations.

Build clients and servers on the Model Context Protocol. You define tools, resources and prompts, connect a client to a server, and review how the protocol handles each exchange.

• Why a protocol for tools and data
• Building a server
• Defining tools
• Defining resources and prompts
• Building a client
• Reviewing message flow

You build A working MCP server and client pair, tested end to end.

Complex tasks call for design. You build workflows that chain steps, route inputs and run work in parallel, compare them with agents, and decide which one fits a task.

• Chaining steps
• Routing inputs to specialists
• Parallelizing independent work
• Where agents differ from workflows
• Choosing the right design
• Measuring each design’s results

You build Three designs for one task, with a comparison of cost, speed and quality.