Plan Optimizer

One pass, not a loop. Rubric → critique → rewrite. The critique and the rewrite are separate acts because judging and generating use different muscles; that separation is where the value is. There is no scoring loop: a model grading its own rewrites in one context will reliably produce a climbing number whether or not the plan improved, so numeric trajectories are theater. The oracle for a plan is the rubric, the user, and the domain docs — not a self-assigned score.

Step 0 — Decide the skeleton (optional, only when the framing is in doubt)

If the plan's basic structure might be wrong — not just its details — generate 3 genuinely different framings first (different architectures, orderings, or theories of the problem, not minor variants). Compare them against the rubric criteria in prose, pick the strongest skeleton, and say in one line why it beat the others. Skip this entirely when the framing is already settled; it's for structural doubt, not ritual.

Step 1 — Build the rubric

Before touching the plan, write the rubric. Derive criteria from what would make this plan succeed or fail, not a generic checklist. Prefer criteria checkable against something objective: does every phase have an exit condition? is there a named first action? is the rollback step present? Most plans are judged on some mix of:

• Goal clarity — is "done" unambiguous and checkable?
• Completeness — is anything load-bearing missing?
• Sequencing & dependencies — right order, prerequisites named?
• Feasibility — realistic given time, people, constraints?
• Risks & mitigations — real failure modes named, with contingencies?
• Specificity — owners, dates, concrete first step?