Examhigh-yieldexam-meta

One-Page Cheat Sheet

Last-hour review. Every fact and reflex on one page.

On this page

One-Page Cheat Sheet

Reflexes (symptom → answer)

See thisAnswer this
Customer wants a humanEscalate (structured handoff summary)
Wrong/inconsistent output formatFew-shot examples
Tool ignored or misusedImprove the tool description
Fabricated parameter valuesDescription: "value must come from prior tool call"
"Must / never / compliance"Hook or in-tool code, never prompt-only
Sources/citations lost between agentsStructured output: content + metadata separated
Context overflow in one agentSubagents explore, return summaries
Slow, incoherent long chatSummarize old turns, keep last 5–6 verbatim
Huge tool responsesTrim to relevant fields
Bulk non-urgent workBatch API (50% off, ≤ 24 h)
Guaranteed JSONTool whose input_schema = target schema + tool_choice (current docs: add strict: true, or Structured Outputs)
Missing data in sourceOptional fields + return null
Edit tool keeps failingRead, then Write
Search file contentsGrep (Glob = file names only)
Files changed on diskResume + say which files changed
Compare 2 approachesFork the session
Stale tool results in old sessionNew session + summary + fresh tool calls
Risky/unfamiliar changePlan mode first
Tiny clear changeDirect execution

Numbers

FactValue
Pass score720 / 1000
Questions / time60 / 120 min
Domain 1 weight (orchestration)27% — the biggest
Cache TTL5 min default, 1 h option
Cache read / write cost0.1× / 1.25× (5-min write) / 2× (1-h write)
Batch API50% off, results ≤ 24 h, match by custom_id
CLAUDE.md import depth4 hops (older material says 5)
Refund hook threshold (canonical example)$500

Flags

FlagOne-liner
--system-promptREPLACES default prompt (loses tool guidance) — the trap
--append-system-prompt(-file)ADDS to default — the fix
--bareNo auto-discovery (no CLAUDE.md, skills, MCP)
-pHeadless; still loads CLAUDE.md unless --bare
--output-formattext / json / stream-json
--max-turns, --max-budget-usdCI cost caps (both print mode only; max-turns exits with an error at the limit)
--dangerously-skip-permissions= bypassPermissions; deny rules still block
--continue / --resume <name> / --fork-sessionLast session / named session / branch

Loop control

  • Continue while stop_reason == "tool_use"; stop at "end_turn".
  • Tool results append to conversation → model reasons about next step.
  • ❌ Don't parse text for "done". ❌ Don't use turn caps as the main stop.
  • Loop ended unresolved? Orchestration code escalates programmatically.

MCP

  • Scopes: local → project (.mcp.json, in git) → user. First wins whole entry.
  • Tools: mcp__server__tool. Prompts: /mcp__server__prompt.
  • Execution failure → isError: true + what to try next.
  • Missing param → protocol error (exam-era answer; current docs prefer isError: true so the model can self-correct).
  • Zero results → normal result, NOT error.
  • Your own errorCategory + isRetryable fields in the result payload (a convention, not MCP spec) fix retry/escalate confusion.

Wrong-answer smells

"automatically" · "unrestricted" · "always/never" · "without review" · "full access" · re-doing work · raw dumps · fixing downstream · prompt-only guarantees · extremes.


Next: Question Anatomy