Claude Certified Architect

1.1 Implement and control agentic loops

• · stop_reason handling (tool_use vs end_turn)
• · tool result handling
• · loop termination — not heuristics or arbitrary caps

1.2 Design multi-agent orchestration patterns

• · Coordinator–subagent hub-and-spoke
• · Task tool spawning
• · Explicit context passing — subagents do not auto-inherit parent memory

1.3 Delegate work to subagents

• · allowedTools must include "Task"
• · Structured handoffs: claim, excerpt, URL, date for synthesis

1.4 Enforce business rules in agent workflows

• · PostToolUse / tool interception hooks
• · Programmatic gates vs prompt-only ordering for refunds and policy

1.5 Normalize heterogeneous tool outputs

• · Consistent date formats, status codes, field naming before agent reasoning

1.6 Task decomposition strategies

• · Prompt chaining for predictable reviews
• · Dynamic decomposition for open-ended investigation
• · Per-file passes + cross-file integration pass

1.7 Session state, resumption, and forking

• · --resume named sessions
• · fork_session for parallel exploration branches
• · Fresh session + summary when tool results are stale

2.1 Design effective tool interfaces

• · Rich descriptions: inputs, outputs, edge cases, when to use vs similar tools
• · Split overlapping tools (e.g. analyze_content → extract_web_results)

2.2 Structured MCP error responses

• · isError flag
• · errorCategory: transient / validation / business / permission
• · isRetryable — avoid wasted retries on policy violations

2.3 Distribute tools across agents

• · 4–5 tools per agent vs 18-tool overload
• · Scoped verify_fact for synthesis
• · tool_choice: auto, any, forced tool

2.4 Integrate MCP into Claude Code

• · Project .mcp.json vs user ~/.claude.json
• · ${ENV} expansion
• · MCP resources for catalogs vs exploratory Grep

2.5 Built-in tools: Read, Write, Edit, Bash, Grep, Glob

• · Grep for content search
• · Glob for path patterns
• · Read + Write when Edit lacks unique anchors

3.1 CLAUDE.md hierarchy and modular rules

• · User vs project vs directory scope
• · @import modular standards
• · .claude/rules/ topic files
• · /memory diagnostics

3.2 Slash commands and skills

• · .claude/commands/ project vs ~/.claude/commands/
• · SKILL.md frontmatter: context: fork, allowed-tools, argument-hint

3.3 Path-specific rules

• · YAML frontmatter paths globs — e.g. **/*.test.tsx regardless of directory

3.4 Plan mode vs direct execution

• · Plan mode for multi-file architecture, monolith→microservices
• · Direct execution for bounded single-file fixes

3.5 Claude Code in CI/CD

• · claude -p for non-interactive runs
• · --output-format json
• · --json-schema for structured CI output

4.1 JSON schemas and tool_use

• · Nullable optional fields to prevent hallucination
• · tool_choice configuration
• · strict schema mode

4.2 Few-shot and review criteria

• · Ambiguous scenario examples
• · Explicit criteria to cut false positives in PR review
• · Multi-pass review architecture

4.3 Validation-retry loops

• · Pydantic / JSON schema failures fed back with document + error
• · Resolvable vs absent-information errors

4.4 Message Batches API tradeoffs

• · 50% cost savings vs latency
• · Blocking pre-merge vs overnight reports
• · custom_id failure handling

5.1 Context window optimization

• · Trim verbose tool outputs
• · Structured fact extraction
• · Lost-in-the-middle — position-aware ordering

5.2 Escalation and human-in-the-loop

• · Policy gaps vs customer-requested escalation
• · Explicit criteria + few-shot — not uncalibrated self-confidence scores

5.3 Multi-agent error propagation

• · Structured failure context to coordinator
• · Partial results + coverage gap annotation

5.4 Information provenance

• · Claim–source mappings
• · Conflicting sources preserved with attribution
• · Temporal data in synthesis

5.5 Human review routing

• · Field-level confidence
• · Stratified sampling
• · Accuracy by document type and field