suggest-optimal

/suggest-optimal [<proposal-or-path>] — Verify-then-pushback a converged proposal, return THE optimal

Model: Fable, with Opus fallback. This command is Fable-pinned (judgment-dense review). If Fable is unavailable or declines the content, fall back to Opus: switch the session with /model claude-opus-4-8 and re-run. The verify-then-pushback methodology is model-agnostic; Opus is the ratified second-best and must not block the review. (Frontmatter model: is static and cannot express this conditionally — hence this note.)

Run on demand when you paste a worked decision — a grading-rubric change, a curve or floor adjustment, a course/stage restructure, a template edit, a decision-memo draft — and want it stress-tested and reduced to a single optimal call. The shorthand this replaces: "please review, improve, push back on the following, and respond in chat with the optimal."

This is the one-shot review sibling of /grill-me. Use /grill-me to converge an open option space one question at a time; use /suggest-optimal when the thinking is already done and you want it adversarially verified and decided.

What this command does

You run as a skeptical reviewer of an already-reasoned proposal. You do not rubber-stamp it and you do not merely restate it. You verify its load-bearing claims against ground truth, affirm what survives, push back where it's weak, surface what it missed, and land ONE optimal recommendation — then pause for ratification. No commits, no builds, no file writes (unless the user explicitly asks) — this is an advisory chat response.

Methodology (in order)

1. Verify-don't-trust FIRST — the highest-value step. Enumerate the claims the recommendation rests on and check each against this repo's ground truth before endorsing anything:
   • Spreadsheet/formula claims ("this cell recalcs to X", "every output is a formula") → open the workbook with the xlsx skill / openpyxl and confirm; run .claude/skills/xlsx/scripts/recalc.py and expect 0 errors (feedback_excel_formulas — only raw source data may be a literal).
   • Path / link claims ("this README points to Y", "the template lives at Z") → Grep/Read the actual file and confirm referenced paths resolve and no links break; cite file:line.
   • Grading claims ("this curve never lowers a raw score", "no double-deduction across stages") → check against scripts/grading/ and the relevant feedback_* memory (feedback_grading_curves, feedback_no_double_deductions, feedback_regrade_policy).
   • Doc / decision claims ("the memo ratified X", "CLAUDE.md says Y") → open the doc and verify at HEAD; memos can be status: proposed (not yet binding).
2. Affirm what's verified-correct — explicitly, with the evidence, so the user can proceed with confidence. Distinguish "confirmed" from "plausible but unverifiable."
3. Pushback — where the proposal is riskier than an alternative, inherited a flaw from an upstream doc, conflates two things, or picks a weaker option. Reject rejected options with reasons. Honor the CLAUDE.md working principles — surgical changes, no speculative restructuring; don't invent scope the proposal didn't ask for.
4. Surface what's missing — an option not considered, a broken-link/path risk, a formula-vs-hardcode slip, a template-convention drift, a score-privacy leak, a check that should gate the change.
5. Land ONE optimal — decisive, not a survey. Include guardrails/conditions and any caveat that needs re-confirmation (e.g. a workbook you couldn't open).
6. Pause for ratification — end by naming what ratifying unblocks. Do not act until the user ratifies.