Learn to build
with Claude.

Start with the ideas behind a model response: how text is represented, how responses vary, and what fits in context.

A reflection to try

Try explaining the difference between a token and a word. Then consider what happens when a conversation outgrows its context window.

Understand how an application talks to an AI model, and where an SDK fits alongside a direct API request.

A reflection to try

Sketch a request traveling from your application to an API and back. Label the request, response, SDK, and error handling.

Begin with the problem, define what success looks like, and follow an AI application from idea to maintenance.

A reflection to try

Choose a small AI project. Write one user need, one measurable acceptance criterion, and one condition that requires human review.

Follow a structured Messages request through images, a read-only tool call, and a continuation that preserves the evidence.

A reflection to try

Sketch a tool request and its matching result. Mark where the application validates inputs, checks access, and decides what action is allowed.

Separate streamed fragments from completed responses, compare interactive work with batches, and plan recovery across provider boundaries.

A reflection to try

Describe a stream that stops before the final event. What can you display, what can you commit to history, and what must you check before retrying?

Trace request contracts, JSON validation, and asynchronous work from an application request to a reviewed result.

A reflection to try

Sketch two independent reads followed by a draft that depends on both. Identify which work may overlap and which steps must wait.

Follow a patch from edited files through staging, commits, merge conflicts and review. Distinguish a refactor from a deliberate behavior change.

A reflection to try

A formatter extraction also changes who may read a repair record. Identify what must remain stable, what behavior changed, and which checks and approval the change needs.

Trace instruction sources across Claude chat, projects, Claude Code and API applications. Keep reference evidence separate from guidance and enforced permissions.

A reflection to try

Move one assistant task between two Claude interfaces. Record which instructions and evidence actually loaded, then explain which controls enforce approval before an action.

Design schema contracts, keep repair-case sessions separate, and choose a maintainable plugin boundary. Includes three original retrieval questions.

A reflection to try

Design a fictional assistant. Name one structural check, one evidence check, and one session boundary. Explain when its reusable procedure needs a standalone skill or a plugin.

Trace effective configuration across model identifiers, API and SDK versions, prompts, settings, and resolved dependencies. Includes three original decision questions.

A reflection to try

A teammate says the prompt is unchanged, so the system is unchanged. Name three other configuration details you would inspect and one observed check that would support your conclusion.

Improve a failing prompt through observable requirements, correct placement, diverse examples, usable constraints, and separate evaluation cases. Includes three original questions.

A reflection to try

Rewrite “summarize professionally” for a fictional workflow. Add a rule and example for missing information, a separate evaluation case, and one requirement enforced outside the prompt.

Preserve decisive evidence and valid state while using targeted retrieval, pruning, compaction, and isolated investigations. Includes three original questions.

A reflection to try

Choose a long-running task. Explain what you would prune, what a summary must retain, and what you would delegate. Write a continuation question that would reveal a lost fact.

Follow an invented delivery record through completion, parsing, structural checks, evidence validation, and authorization. Practice three original decisions about missing fields, truncation, and permission checks.

A reflection to try

Walk an invented delivery record through completion, parsing, structure, and evidence checks. Explain which failures need correction, a bounded retry, or human review.

Compare model configurations by acceptable-result quality, end-to-end latency, compatibility, and total cost. Use fictional token prices to calculate retries and cache reuse, then connect configuration, prompting, context, and output handling in a cumulative checkpoint.

A reflection to try

Compare two fictional workloads. Define acceptable results, a latency measure, and a token budget. Explain when cache reuse would save money and what evidence would change your model choice.

Design a predictable packing workflow and an adaptive investigation for a fictional museum. Choose routing, parallel checks, single-agent investigation, or supervised delegation, and define concrete completion, limits, and human handoffs.

A reflection to try

Sketch a predictable workflow and an evidence-dependent investigator. Identify who chooses each step, what can run independently, who reconciles findings, and two stopping conditions.

Trace a tool-use loop, distinguish active context, checkpoints, and long-term memory, and compare framework abstractions. Use original scenarios to reason about denied tools, scoped preferences, durable review pauses, and recovery evidence.

A reflection to try

Design the smallest useful assistant for a fictional workshop. Distinguish active context, a checkpoint, and long-term memory, then name one framework capability you would verify.

Compare a custom Messages API loop, a process-operated Claude Agent SDK harness, and Claude Managed Agents. Follow Maple Room through execution ownership, conversation and working state, self-hosted tool sandboxes, hooks, and recovery after an uncertain write. Includes original retrieval questions and an ownership teach-back.

A reflection to try

Compare a custom API worker, an Agent SDK worker, and a hosted harness with self-hosted tools. Name who runs the loop, executes tools, stores state, and verifies recovery, including which data travels.

Follow a proposed repair-note operation through schema design, registered dispatch, business validation, access checks, approval, execution, and matched results. Original questions cover forbidden but well-formed inputs, parallel lookup results, and recovery after a write timeout.

A reflection to try

Trace a proposed repair note through dispatch, validation, access, approval, execution, and its matched result. Explain how you would handle an unauthorized record, refused approval, and lost acknowledgment.

A compact MCP lesson: identify host, client, and server; distinguish resources, prompts, and tools; select compatible protocol and transport behavior; and test clear errors and denied access. Current2026-07-28 request semantics are explicitly separated from older initialization-based revisions.

A reflection to try

Sketch an MCP host, client, and server for a fictional repair workflow. Distinguish a manual resource from a lookup tool, then identify the transport and compatible protocol revision.

Build Maple Room's handover workflow by matching existing operations, a custom repair integration, reusable Skill guidance, an MCP connection, and optional plugin packaging. Learn why these components can compose, how execution and permissions differ by surface, and which maintenance responsibilities each introduces.

A reflection to try

Separate a workflow’s existing capabilities, missing business operation, and reusable procedure. Justify whether MCP or a plugin adds enough value, and name one component you would leave out.

Choose Claude Code components, inspect loaded project guidance, resume sessions carefully, and distinguish output modes from permission and isolation controls.

A reflection to try

Resume a fictional coding task with a missing tool connection and an interrupted ticket request. Name the evidence needed to continue, and distinguish conversation state, permissions, and isolation.