Planner-Worker Model Tiering

Right instinct — the burn is in implementation, not planning. This wraps
your build in a planner-first flow that keeps the premium model OUT of
the code entirely.

THE FLOW
  1. PLANNER (premium model) reads the PRD, asks clarifying questions,
     produces a spec-driven multi-phase architecture. This is where
     judgment matters and it's cheap — planning is a fraction of a
     build's tokens.
  2. Each phase hands off to a CHEAPER sub-agent, matched to the work:
       Research / exploration   → Haiku (cheapest, fine for reading)
       Scaffolding / boilerplate → Sonnet
       Heavier implementation    → Opus 4.8
  3. AUTOMATIC CHECKPOINTS + self-review between phases, so a cheap
     worker going off-spec is caught before it compounds.

THE RULE THAT SAVES THE TOKENS
The premium model NEVER touches implementation. It plans, it reviews at
checkpoints, and it delegates. Implementation — the token-heavy
back-and-forth you're worried about — runs on cheaper models. That's the
whole savings mechanism: you're paying premium rates for judgment, cheap
rates for typing.

WHY CHECKPOINTS MATTER HERE
A cheap worker with no oversight drifts, and fixing drift costs more than
it saved. The checkpoints let the planner catch a phase that went sideways
and correct course — so you get cheap-model economics with premium-model
judgment on the decisions that matter.

WHAT YOU GET
A long build that would have burned your weekly limit runs mostly on
sub-agents, with the premium model spending its tokens on the plan and the
checkpoints — the two places its judgment actually changes the outcome.

Point me at the PRD and I'll run the planner pass first: clarifying
questions, then the phased architecture with a model assigned per phase.