offload-to-locallisted

Use when PLANNING or decomposing ANY multi-step task, and whenever your plan involves reading, searching, summarizing, transforming, drafting, extracting, or editing across files — regardless of whether cost is mentioned. Also when you're about to do delegatable/bulk/mechanical work on a paid cloud model, or you notice you're about to burn cloud tokens on work a smaller model could do. It routes delegatable work to a free local model — by DEFAULT to the broad `operator` role, escalating to a specialist role only when needed — instead of spending cloud tokens. Explicit triggers: starting a multi-step plan (especially under a budget constraint / Credit Efficient Mode), "save cost/tokens", "do this locally", "offload this", "use the local model for this". Do NOT use it to offload frontier-reasoning, architecture, security-critical, or final-review work.
haiggoh/local-agents · ★ 1 · AI & Automation · score 74

Install: claude install-skill haiggoh/local-agents

# offload-to-local — send the legwork to a free local model The biggest cost saving in Claude Code is not a cheaper cloud model — it's **$0 local compute**. When your main session runs a capable cloud model (Opus/Sonnet), keep that model for judgment and **delegate the bulk/mechanical legwork to a local model** via dispatch. This applies to anyone who wants to save cost, with or without a spending cap. ## Decide UPFRONT, at decomposition — not mid-work The habit is not "notice you're already grinding through files and stop." It's: **when you first break a task into steps, decide per step where it runs before you execute anything.** Waiting until you're mid-bulk-work means the cloud tokens are already being spent. **Decide before you READ the inputs.** This is the sharpest version of the rule, and the easiest to get wrong. If you open the files to judge whether the task is worth delegating, you have already done the expensive part — and delegating afterwards is theatre, because the work is spent either way. So "I've already read the material, it'd be quicker to just do it now" is **evidence the decision came too late**, not a reason to keep it on cloud. Decide from the *shape* of the step, which you know at decomposition, not from its contents. If you catch yourself there anyway, a dispatch still buys something real: run it as an **independent cross-check** of the work you just did. Diffing your output against a local model's is a genuine second use, and it costs nothing.