credit-routinglisted

Use at the start of any non-trivial task, when deciding whether to delegate, when a task has failed twice, before escalating model or effort, or when the user mentions credits, cost, tokens, routing, or 'which model'.
courtneyr-dev/agent-skills · ★ 0 · AI & Automation · score 72

Install: claude install-skill courtneyr-dev/agent-skills

# Credit routing Claude Code is the orchestrator, reviewer, and final judge — not the typist. Spend expensive tokens on judgment; route mechanical work to cheaper executors; make knowledge compound instead of re-deriving it. Canonical references (read just in time, not preemptively): - Routing matrix + examples: `~/.claude/knowledge/wiki/tools/tool-routing.md` - Model/effort policy + failure diagnosis: `~/.claude/knowledge/wiki/tools/model-effort-policy.md` - Delegation prompt template: `references/delegation-template.md` - Verification, explainers, compaction: `references/verification-checklist.md` - Unknowns mapping + context rules: `references/unknowns-and-context.md` - Benchmark log + scorecard: `~/.claude/knowledge/wiki/tools/benchmark-log.md` ## The loop (every non-trivial task) 1. **Observe** — inspect only the needed files and current state. Grep/glob/index before opening files; Explore subagent for bulk sweeps. 2. **Orient** — check the wiki index (`~/.claude/knowledge/wiki/index.md`) and project instructions before re-deriving anything. For unfamiliar high-impact work, run a blind spot pass first (see unknowns-and-context.md). 3. **Decide** — route per tool-routing.md: Claude keeps judgment; subagents take bounded analysis; Gemini takes second-pass reading/critique; OpenClaw takes bounded personal automation; scripts take anything deterministic; ask the user only when the answer changes the architecture. 4. **Act** — smallest effective unit of work. Delegations