Quick reference tables

Debugging & logging

Situation	Recommended action
Claude does something wrong	/log_error -> fork -> interview -> capture verbatim prompt -> rewind
Claude does something right	/log_success -> capture prompt -> extract template -> consider promoting to CLAUDE.md
Same error keeps recurring	Review error logs -> identify pattern -> create hook or add to CLAUDE.md
Claude hallucinating APIs	Use Context7 MCP for current documentation -> category: hallucination

Workflow execution

Situation	Recommended action
Need reliable workflow execution	Use /command with deterministic wrapping, not natural language
Repetitive multi-step task	Create command in .claude/commands/ -> document purpose + workflow
Different phases need different models	Implement model routing in command definition
Tasks can run in parallel	Design command to spawn independent subagents -> maintain isolated contexts

Permission & context management

Situation	Recommended action
Claude requesting excessive permissions	Implement hooks + --dangerously-skip-permissions for safe autonomy
Need to block dangerous operations	Create pre-execution hooks in .claude/hooks/ -> pattern match blocked commands
Context getting bloated	Use /compact at natural boundaries -> consider subagent delegation
CLAUDE.md exceeds 50 lines	Audit weekly -> move rarely-used instructions to skills -> keep only essentials
Switching between projects	/clear with repo-specific config -> clean context transition

Recovery & restoration

Situation	Recommended action
Bugs polluted the context	Double-escape -> restore conversation only (preserves correct code)
Need full rollback	Double-escape -> full restore (reverts conversation + code)
Claude going in circles	Fork conversation -> rephrase approach -> try different angle
Ending session cleanly	/handoff {NOTES} -> explicit priority notes for next session

Multi-agent patterns

Situation	Recommended action
Complex reasoning task	Force Opus subagent -> add "Always launch opus subagents" to config
Subagent producing poor results	Check model selection -> likely model mismatch -> upgrade to Opus
Errors compounding across agents	Implement deterministic verification between handoffs -> don't trust claims
Subagent handling too many tasks	Split into multiple focused agents -> one objective per agent

Input optimization

Situation	Recommended action
Typing prompts feels slow	Use reprompter tool -> voice input -> automated structure -> keybind activation
Unsure what to ask for	Model interview approach -> "Interview me about what I want" -> let model surface requirements
Prompt too vague	Add XML structure -> role assignment -> specific examples -> clear constraints
Same prompt type needed repeatedly	Extract template -> store in success logs -> promote to skill if valuable