Quick Reference Cards

AGENTIC LOOP: stop_reason == "tool_use" → execute tools, append results, continue stop_reason == "end_turn" → done ANTI-PATTERNS: ✗ Parse natural language ("I'm done") ✗ Arbitrary iteration caps as primary mechanism ✗ Check content[0].type == "text" MULTI-AGENT: ✓ Hub-and-spoke (coordinator at centre) ✓ Subagents are isolated (no shared memory, no inherited history) ✓ All communication through coordinator ✓ Trace failures to root cause (usually coordinator decomposition) ENFORCEMENT: High stakes (financial/security/compliance) → Programmatic (hooks/gates) Low stakes (style/format) → Prompt-based guidance HOOKS: PostToolUse → Normalise data after tool execution Pre-execution → Block/redirect before tool execution DECOMPOSITION: Fixed pipeline → Predictable tasks Dynamic adaptive → Open-ended investigation Attention dilution → Split into per-file + cross-file passes SESSION: Resume → Context still valid Fork → Divergent exploration Fresh+summary → Stale context or degraded session

TOOL DESCRIPTIONS: Include: purpose, inputs, examples, edge cases, boundaries Fix misrouting with better descriptions FIRST Check system prompt for keyword conflicts ERROR CATEGORIES: Transient → retry (timeout, unavailable) Validation → fix + retry (bad input) Business → alt workflow (policy violation) Permission → escalate (access denied) Access failure ≠ valid empty result TOOL DISTRIBUTION: 4-5 tools per agent, scoped to role tool_choice: auto / any / specific tool Scoped cross-role tools for high-frequency simple ops MCP: .mcp.json = project (shared, version-controlled) ~/.claude.json = user (personal) Community servers before custom builds ${ENV_VAR} for credentials BUILT-IN TOOLS: Grep = search contents Glob = match file paths Edit = targeted modification Read + Write = fallback when Edit fails

CLAUDE.md HIERARCHY: ~/.claude/CLAUDE.md → personal, NOT shared .claude/CLAUDE.md → project, shared via git subdirectory/CLAUDE.md → directory-scoped New hire missing instructions? → Probably in user-level config PATH-SPECIFIC RULES: .claude/rules/testing.md with paths: ["**/*.test.tsx"] Loads only when editing matching files Glob patterns match across entire codebase COMMANDS & SKILLS: .claude/commands/ → project (shared) ~/.claude/commands/ → personal Skills: context: fork, allowed-tools, argument-hint Skills = on-demand | CLAUDE.md = always-loaded PLAN MODE: Complex/multi-file/architectural → Plan mode Simple/clear/single-file → Direct execution Common: Plan first, then direct execution ITERATIVE REFINEMENT: Input/output examples > prose descriptions Batch feedback when fixes interact Sequential when issues are independent CI/CD: -p flag = non-interactive mode (MEMORISE THIS) --output-format json for structured output Independent review instance (not same session) Include prior findings to prevent duplicate comments

EXPLICIT CRITERIA: Specific categories > vague confidence instructions Code examples for severity levels Disable high false-positive categories temporarily FEW-SHOT: Most effective for consistency 2-4 examples with reasoning Deploy when prose produces inconsistent results STRUCTURED OUTPUT: tool_use eliminates syntax errors, NOT semantic errors Nullable fields prevent fabrication "unclear" enum for ambiguous cases "other" + detail string for extensibility VALIDATION-RETRY: Effective: format, structural, placement errors Ineffective: information absent from source Send: original doc + failed extraction + specific error BATCH API: 50% cost savings, up to 24h, no latency SLA Blocking workflows → synchronous Latency-tolerant → batch No multi-turn tool calling in batch MULTI-INSTANCE: Same session = biased self-review Independent instance = objective review Per-file passes + cross-file integration pass Confidence routing → human review for low-confidence

CONTEXT PRESERVATION: Persistent case facts block → never summarise transactional data Key findings at beginning (lost-in-middle effect) Trim tool results to relevant fields Full conversation history in every API request ESCALATION: Customer says "I want a human" → immediate, no investigation Policy gap → escalate Frustration alone → NOT a valid trigger Self-reported confidence → NOT reliable ERROR PROPAGATION: Include: failure type, what was attempted, partial results Never: silent suppression or full workflow termination Access failure ≠ valid empty result CODEBASE EXPLORATION: Scratchpad files for persistent findings Subagent delegation for investigations /compact when context is bloated Crash recovery via structured manifests CONFIDENCE CALIBRATION: Validate by document type AND field, not just aggregate Stratified random sampling for ongoing verification Field-level confidence → route low-confidence to humans PROVENANCE: Structured claim-source mappings through entire pipeline Conflicting sources → annotate both, let consumer decide Temporal differences ≠ contradictions Content-appropriate rendering (tables, prose, lists)