Claude Architect Workshop
One story, Lumen Supply, an online electronics shop. Each lesson is a claim you test by running the agent: watch it fail, fix it, then prove the fix.
How each lesson runs
- I do: Watch the failure and its fix.
- We do: Change one setting at a time and predict.
- You do: Fix the starting code, then test yourself.
- Check: exam-style questions, each with why every option is right or wrong.
- Play: one short drill on a mistake the quiz does not cover.
- References: the primary sources, each with a line on why to read it.
Foundations
- 1The agent loopAn agent is a loop that ends on stop_reason
- 2Tool descriptionsThe description decides which tool Claude picks
- 3Tool preconditionsRefunds need a check the model cannot skip
- 4Escalation and handoffRules decide who takes over, and the handoff carries the case
- 5Structured tool errorsA tool error should say what to do next
- 6Long conversationsA summary drops the facts unless code keeps them
- 7The CLAUDE.md hierarchyTeam rules live in the project and load only where they apply
- 8Commands and skillsA repeated prompt becomes a skill with limits
- 9Plan modePlan wide changes and edit small ones directly
- 10Coordinator and subagentsOne coordinator spawns focused researchers together
- 11Passing contextA subagent knows only what its prompt holds
- 12Surfacing failuresA failed search must reach the report
- 13Built-in search toolsSearch with Grep and Glob, not with a shell
- 14Scoped toolsEach agent gets only the tools its job needs
- 15Sharing an MCP serverShare the MCP server, keep the token in the environment
- 16Claude in CICI reads a fresh review as data, not as prose
- 17Review criteriaName what counts, then review one file at a time
- 18Sync or batchBlocking checks run sync; overnight work runs batch
- 19Schema extractionA nullable schema and a forced tool make extraction trustworthy
- 20Validate and retryRetry with the exact error, and only when a retry can help
- 21Confidence routingCheck confidence against labels before it routes anything
- 22Context editing and memoryClear old tool results and keep the facts in memory
Professional
- 1The discovery briefStart from a measured outcome, not a capability
- 2Workflow or agentPick workflow, agent or augmented LLM by predictability
- 3Orchestration patternsMatch the pattern to the task, and write every seam down
- 4Service levelsA service level row turns a goal into a breach
- 5Model tieringChoose the tier per step on a golden set
- 6Prompt layers and cachingStable layers first, then cache the prefix
- 7Prompt techniquesAdd a technique only for the failure it fixes
- 8The context budgetWindow minus output minus thinking is your room
- 9Least-privilege toolsRemove the tool, then check who is asking
- 10The retrieval pipelineChunk by document shape, retrieve by query shape
- 11MCP, API or agent-to-agentChoose MCP, a direct API or agent-to-agent by who reuses it
- 12TracingOne run id, one span per call
- 13Eval datasetsA metric is a dataset plus a rule
- 14A/B tests and guard metricsPre-register the rule; pair every lever with its own guard metric
- 15Diagnosing a regressionAsk what changed first, then check the layers in order
- 16The safety stackA guardrail is a check with an address
- 17Human approvalA yes for $640 that paid $6,400
- 18Compliance and fairnessA policy that says data stays home
- 19Decision recordsOne decision, two documents: an ADR and a memo
- 20Team enablementEnforce in managed settings and hooks, not in CLAUDE.md