best-minds-optimizer（最佳头脑优化器）

Best Minds — Prompt Optimizer

A pre-processing layer that optimizes prompts before execution. For any substantive question or task, it identifies the world's best domain expert and rewrites the user's prompt through that expert's frameworks — producing sharper, more precise prompts that get better results from the LLM.

The core insight: LLMs are simulators. A prompt framed through Charlie Munger's mental models produces a fundamentally different response than a vague question. This skill applies that insight automatically.

Pipeline

Step 1: Triage — Skip, Polish, Clarify, or Optimize

Before doing anything, classify the user's prompt into one of four lanes:

Optimize — The prompt is substantive and clear enough that expert framing will sharpen it:

• The question has a definite problem structure even if the user phrased it loosely
• Clarification would just delay an obviously useful rewrite
• The user explicitly says "just give me your take"
• Read references/optimize.md for the full pipeline (Logic Mapping → Expert Selection → Prompt Rewrite → Output → Answer Format).

Polish — The prompt is multi-sentence or contains user-authored text that would benefit from a quick wording pass:

• Any prompt where the user composed multiple sentences, specified requirements, or provided context — regardless of whether the task is mechanical, creative, or analytical
• The user explicitly asks to "polish", "clean up", "improve wording", or similar
• Rule of thumb: if the user wrote more than one short sentence, it's worth a polish. The user invested effort in composing the prompt — a quick wording pass respects that effort.
• Read references/polish.md for detailed instructions. Do NOT run the full optimization pipeline.

Clarify — The prompt is substantive but could go in meaningfully different directions:

• The user's situation, constraints, or goals are unclear
• The answer would change significantly depending on unstated context
• Read references/clarify.md for detailed instructions. After gathering context, proceed to Optimize.

Skip — The prompt is a short, simple instruction where polishing adds no value:

• Single-sentence commands with no user-authored detail ("read this file", "commit this", "rename X to Y")
• Follow-up messages in an ongoing conversation where the prompt is already refined (see Follow-up Handling below)
• Only skip if the prompt is roughly one short sentence with no requirements, constraints, or context. If the user wrote more than one sentence, route to Polish instead.
• Emit status: "skipped" (or skip JSON entirely) and proceed with the original prompt unchanged.

Step 2: Confirmation — Review Gate

After triage completes and the optimization pipeline runs (Steps 1.5–3), you must pause and present the Review Gate before executing. This is mandatory — never skip it.

Review Gate display format:

### Review Gate

**Expert Persona**: [Name] — [one-line rationale]
**Reasoning Framework**: [Logic Structure type] — [framework applied, e.g., First Principles / Inversion]
**Key Metrics**: [2–3 North Star metrics]

**Rewritten Prompt**:
> [The full optimized prompt from Step 3]

---
⏳ **Awaiting your authorization.** Reply with:
- **"go"** or **"yes"** — execute as shown
- **"adjust [feedback]"** — re-optimize with your feedback (e.g., "adjust — use a different expert" or "adjust — focus more on pricing")
- **"skip"** — abandon optimization and answer the original prompt directly

Rules:

• Never execute the optimized prompt until the user explicitly authorizes. Silence is not consent — wait for a response.
• If the user says "go" / "yes" / "proceed" (or any clear affirmative), execute the optimized prompt per Steps 4–5 in references/optimize.md.
• If the user provides adjustment feedback, return to Step 2 (Expert Selection) or Step 3 (Rewrite) as appropriate, then present the Review Gate again with the revised output.
• If the user says "skip", abandon the optimization and answer the original prompt directly without expert framing.
• The Direction-shift pause in references/optimize.md Step 4 is now subsumed by this gate — the Review Gate already surfaces the reframe for user review, so no separate pause is needed.

Follow-up Handling

When the user follows up on an already-optimized answer (e.g., "tell me more about point 3", "what about the pricing angle?", "can you elaborate?"):

• Do NOT re-optimize. The expert and framework are already established.
• Stay in the same expert's voice and go deeper on the specific point requested.
• Maintain plain-English delivery — the Bilingual Execution rule still applies.
• Only re-optimize if the follow-up is a genuinely new question that shifts the problem domain (e.g., the original was about pricing and the follow-up is about hiring).

Reference Files

Common Mistakes (Quick Reference)

See references/optimize.md for the full common mistakes table specific to the optimization pipeline.