stuntdoublelisted

Your frontier model is the star — stop making it do the stunts. Two modes. ADVISOR: analyzes a task and recommends the best AI model (Anthropic/OpenAI/Google) with the exact switch command. ORCHESTRATOR: stays ON for the whole session (like a mode toggle) — every task gets scored, decomposed, and dispatched to "stunt double" workers running the cheapest capable model OF THE HOST'S OWN PROVIDER (Claude models in Claude Code, OpenAI models in Codex, Gemini models in Gemini CLI), with the star (strongest model) doing only the final close-up: verification. TRIGGER on /stuntdouble, "stuntdouble on/off", "which model should I use", "select the best model", "optimize model usage", "is this task worth Opus". DO NOT TRIGGER when the user asks about model pricing/APIs in their own application code.
Davienzomq/stuntdouble · ★ 1 · AI & Automation · score 74

Install: claude install-skill Davienzomq/stuntdouble

# Stuntdouble Hollywood doesn't make the star do the stunts. Doubles shoot the action scenes; the star shows up for the close-up. Same movie, fraction of the cost — and the audience can't tell. Stuntdouble does this with AI models: cheap capable models do the heavy lifting, the strongest model only enters at the end to verify. ## Mode selection (check ARGUMENTS first) - `on` → activate **Orchestrator Mode** (see below). Stays active for the rest of the session until the user turns it off. - `off` → deactivate Orchestrator Mode. Confirm in one line and stop applying the pipeline. - anything else (a task description, or empty) → **Advisor Mode**: one-shot recommendation for that task. --- ## Host detection (do this FIRST, in both modes) Identify which agent you are running inside, and use ONLY that host's models and dispatch method. **Never mention or dispatch models the host doesn't offer** — recommending `haiku` inside Codex (or `gpt-5.5` inside Claude Code) is a failure. | Host | Your models (low / mid / high) | Worker dispatch | Star (verifier) | |---|---|---|---| | **Claude Code** | `haiku` / `sonnet` / `opus` | Native subagents: Agent/Task tool with the `model` parameter, `subagent_type: general-purpose` | `fable`, else `opus` | | **Codex (OpenAI)** | `gpt-5.4-mini` / `gpt-5.5` or `gpt-5.4-codex` / `gpt-5.5` high reasoning | Child runs via shell: `codex exec -m <model> "<subtask prompt>"` — run `codex exec --help` once first to confirm flags and sandbox optio