Get the free Advisor Tool Playbook
The companion to the video, every code identifier verified against Anthropic's docs. THE IDEA: a fast, low-cost EXECUTOR model (Sonnet or Haiku) consults a higher-intelligence ADVISOR model (Opus) mid-generation — the advisor reads your full transcript, hands back a short plan, and the executor keeps generating. You get near-top-model quality at near-cheap-model prices on long agentic tasks. It's beta: send the header 'anthropic-beta: advisor-tool-2026-03-01' (available on the Claude API and Claude Platform on AWS). THE ENTIRE SETUP is one entry in your tools array — type 'advisor_20260301', name 'advisor', model 'claude-opus-4-8', plus 'max_tokens: 2048' to cap it. The executor emits a server_tool_use with an empty input:{} (it only signals WHEN); the server forwards the full transcript to the advisor server-side; the advisor thinks with no tools (its thinking is dropped) and only the advice returns — all inside ONE /v1/messages call. THE PAIRING RULE: the advisor must be Sonnet 4.6 or stronger AND at least as capable as the executor (invalid pairs return a 400); the playbook has the full who-can-advise-whom table. TIMING IS EVERYTHING — call the advisor at two moments: (1) EARLY, after a little orientation but before you commit to an approach, and (2) FINAL, right before you declare the task done; the sweet spot is 2–3 calls per hard task. THE NUDGE: Haiku is shy — a one-line reminder after turn 1 lifts its success ~7 points; the same nudge does nothing for Sonnet and makes Opus slightly worse, so nudge the junior and trust the senior. KEEP IT CHEAP with three settings: cap it (max_tokens 2048 cuts advisor output ~7x), ask for brevity (a 'keep it under 80 words' line in your user message), and cache long runs (breaks even at ~3 calls). Plus a paste-ready system-prompt block for coding executors, the full error-code table (the request never fails, it just continues), and how advisor tokens bill separately under usage.iterations. When to use it: coding agents, computer use, multi-step research. When to skip it: single-turn Q&A, simple lookups, or work where every step needs the big model anyway.
Free. No spam. Unsubscribe anytime.