Leverage is the discipline of multiplication: running work in parallel, isolating agents in worktrees, packaging repeatable playbooks as skills, orchestrating fleets, running unattended overnight, and routing each task to the cheapest model that can do it well.
When tasks are independent and verifiable per-unit: migrations across many files, audits across many pages, research across many sources. Parallelism does not fix a task you cannot verify — it gives you the same doubt at ten times the speed.
A skill is a playbook you wrote once for work you do repeatedly: the deploy steps, the review checklist, the report format. The third time you type the same instructions is the signal. Skills turn your best run into your default run.
Unattended work needs pre-decided boundaries: what the agent may do alone, what it must queue for a human, and where it writes its trail. The overnight failure mode is not wrong code — it is unreviewable code with no record of the decisions.
Match the tool to the stakes. Frontier models for architecture, judgment, and hard debugging; fast cheap models for mechanical edits and summaries. Routing everything to the biggest model burns budget; routing everything to the smallest burns trust.
You have 40 near-identical config files to migrate and one hour. What is the highest-leverage approach?
C. C proves the pattern first, then multiplies it with verification per unit — the core leverage loop. A spends the hour serially, B spends it manually, and D produces an unreviewable blob.