EN / 中文
5Q  ·  五行煉化提示詞診斷
5Q  ·  Five-Quality Prompt Diagnostic

你問的問題,
決定了你能得到的世界

The question you ask,
determines the world you receive

The shape of your question
determines the depth of your answer.

五德 · 界 象 質 律 流

大多數人以為 AI 回答得不好,是模型的問題
5Q 說:說不定是你問問題的方式不對。

調整的不是模型。調整的可能是你。

Most people assume bad AI responses are a model problem.
5Q says: maybe it's how you ask.

You don't adjust the model. You might need to adjust yourself.

「同樣的 AI,為什麼有人能聊出深度,
有人只得到 out of Token ——and a bill coming?」
"Same AI. Why does one person get depth, while another gets out of Tokens — and a bill?"
五德 Five Qualities
核心洞察 · Core Insights
Core Insights · 核心洞察

為什麼提示詞的品質,
比模型的能力更重要

Why the quality of your prompt
matters more than the model's capability

01

AI 的輸出品質,上限由你的輸入決定

Your input quality sets the ceiling for AI output

模型沒有辦法猜出你真正想要什麼。 模糊的問題,得到的是模糊的答案—— 不是因為 AI 不夠好, 而是因為你的需求還沒有成形。 5Q 的第一步,是讓你看見自己問題的形狀。

The model cannot guess what you truly want. Vague questions yield vague answers — not because AI isn't capable, but because your intent hasn't taken shape yet. 5Q's first step is to make the shape of your question visible.

02

這不是 Prompt 技巧,這是認知診斷

This isn't prompt technique. It's cognitive diagnosis.

市場上的 prompt 工具教你「怎麼寫」—— chain-of-thought、few-shot、system prompt。 這些都是。 5Q 問的是更深的問題: 「為什麼你的思路本來就不清楚?」 技巧可以複製,診斷需要對照自己。

Prompt tools teach you how to write — chain-of-thought, few-shot, system prompts. These are craft. 5Q asks the deeper question: "Why isn't your thinking clear to begin with?" Techniques can be copied; diagnosis requires looking at yourself.

03

用久了,你不再需要它——這才是成功

When you no longer need it — that's the goal.

好的老師的目標,是讓學生不再需要自己。 5Q 不是要讓你依賴一個工具, 而是讓你內化五德的思維方式, 讓「問出好問題」成為你的本能, 而不是每次都要查攻略。

A good teacher's aim is to make themselves unnecessary. 5Q isn't about depending on a tool — it's about internalizing the Five Qualities framework, so that asking good questions becomes instinct, not something you look up each time.

五德框架 · The Five Qualities
The Five Qualities · 五德框架

任何存在實體,同時具備五種內在維度

Any entity simultaneously holds five inner dimensions

界 · Boundary

邊界確立。你知道自己在問什麼,也知道不在問什麼。問題無界,能量四散;界定清楚,AI 才能聚焦。

Boundary — Edges defined. You know what you're asking, and what you're not. Unbounded questions scatter energy; a clear boundary lets AI focus.

象 · Manifestation

意圖顯化。你要達成的目的清楚呈現——不是讓 AI 猜你需要什麼格式、深度或用途。

Manifestation — Intent made explicit. Your goal is present and visible — AI needn't guess your desired format, depth, or purpose.

質 · Substrate

素材承載。背景脈絡給足,AI 才能站在你的位置思考,而不是從零開始瞎猜。素材不足,規則無從淘煉。

Substrate — Sufficient grounding. Context and background give AI a footing to reason from your position. Sparse substrate means rules cannot emerge.

律 · Regulation

規則立定。格式、標準、評判方式事先確立,收斂而不失控。精煉不是刪字,是讓規則從問題中自然生出。

Regulation — Rules established. Format, standards, and criteria set upfront — convergent, not chaotic. Refinement isn't cutting words; it's letting structure arise naturally.

流 · Circulation

脈絡流通。問題的來龍去脈貫通,AI 才能接上你的思路,而不是在斷點上重新起跑。

Circulation — Connected flow. The thread of your question runs through — AI follows your reasoning rather than restarting at every gap.

工具示範 · Live Output
Live Output · 工具示範

它跑起來長這樣

What it looks like when it runs

5q — pro · jina-v3
$ ./run.sh --bin onnx --file data/test_prompt.md --model-dir models/jina-v3 --style pro [5Q] 偵測到 NVIDIA GPU:CudaGpu { vram_mb: 6144, device_id: 0 } [5Q] 選用:models/jina-v3/model_fp16.onnx (F16 precision) [5Q] 載入 ORT 模型:models/jina-v3/model_fp16.onnx ✅ cuDNN dylib 預載完成   ✅ preload libonnxruntime_providers_shared.so ✅ 已向 ORT 註冊 CUDA 執行提供者(顯存無上限,按需分配) [5Q] 模型加載完成,準備進行意象推演...   載入 Tokenizer / 模型家族:Jina ╔══════════════════════════════════════╗ 🔍 5Q 提示詞架構診斷報告(專業版) ╚══════════════════════════════════════╝ Token 長度(字元數):481 chars ── [語系校準] 語意基準錨點 ────────────────────────────── 語系基準:zh 靜態權重:38.1966% 語意權重:61.8034% ── [萃取] 指令特徵識別 ── 定界 / — Boundary ██████████ 20/20 優良 顯化 / — Manifestation ████████░░ 16/20 優良 承載 / — Substrate ████░░░░░░ 9/20 不足 ⚠ 定律 / — Regulation ███████░░░ 14/20 達標 流通 / — Circulation ███████░░░ 14/20 達標 ── 綜合評估結果 ──────────────────────────────────────────── total ███████░░░ 73/100 🔴 高風險(易產生幻覺或失控輸出) ── [建議] 優化與修復方向 ─────────────────────── [ · 建議] 補充充分的背景知識、前置條件或 Few-Shot 範例, 確保模型基於事實生成,而非無根據推測。 ── [評估] 多維度語意與結構分析 ── 定界 / — Boundary ████████░░ 16/20 優良 顯化 / — Manifestation ███████░░░ 14/20 達標 承載 / — Substrate ████░░░░░░ 8/20 不足 ⚠ 定律 / — Regulation ███████░░░ 14/20 達標 流通 / — Circulation ███████░░░ 13/20 達標 ── [規模估算] Token Cost 預估 ───────────── GPT-4o → $0.35 (單次) Claude Sonnet → $0.52 (單次) DeepSeek → $0.007 (單次) ⏱ 效能報告 (Timing) ├─ 關鍵字掃描: 145.95ms ├─ 模型載入: 606.03ms ├─ 語意推論: 4.02s └─ 總耗時: 4.17s $ _

真實輸出 · 大型文件 52201 chars · Daemon 離線自動 fallback 本地推論 · vegan 語言包特化

5q — ori · jina-v3
$ ./run.sh --bin onnx --file data/test_prompt.md --model-dir models/jina-v3 --style ori [5Q] 偵測到 NVIDIA GPU:CudaGpu { vram_mb: 6144, device_id: 0 } [5Q] 選用:models/jina-v3/model_fp16.onnx (F16 precision) ✅ cuDNN dylib 預載完成   ✅ preload libonnxruntime_providers_shared.so ╔══════════════════════════════════════╗ 🔮 5Q 提示詞診斷報告 🐕 ╚══════════════════════════════════════╝ 字數(Char count):481 🔷 ── [偵測] 木火土金水能量分布 🐾 ──────────────── 定界 / 🌳 界 ██████████ 20/20 ✨ 優秀 顯化 / 🔥 象 ████████░░ 16/20 ✨ 優秀 承載 / 🏔️ 質 ████░░░░░░ 9/20 ❌ 不足 定律 / ⚙️ 律 ███████░░░ 14/20 ⚠️ 達標 流通 / 🌊 流 ███████░░░ 14/20 ⚠️ 達標 ── 綜合評分 ───────────────────────────────────────────── total ███████░░░ 73/100 🚨 高風險 (可能幻覺) 🐕 建議 多給一些背景資料、Few-Shot、或前置條件, 讓毛孩不要憑空瞎猜噢 🦴 ⏱ 效能 └─ 總耗時: 4.17s $ _

馬爾濟斯版(ori):適合非專業場景 · 簡化訊息 · 表情符號視覺引導

5q — vegan · jina-v3
$ ./run.sh --bin onnx --file data/test_prompt.md --model-dir models/jina-v3 --style vegan [5Q] Detected NVIDIA GPU: CudaGpu { vram_mb: 6144, device_id: 0 } [5Q] Selected: models/jina-v3/model_fp16.onnx (F16 precision) ✅ cuDNN dylib preloaded   ✅ ONNX Runtime providers registered ╔═══════════════════════════════════════╗ 🌿 5Q Prompt Diagnostic Report ╚═══════════════════════════════════════╝ Token length (char count): 481 chars ── [Extract] Instructional Features ── Boundary ██████████ 20/20 Excellent Manifestation ████████░░ 16/20 Excellent Substrate ████░░░░░░ 9/20 Insufficient ⚠ Regulation ███████░░░ 14/20 Adequate Circulation ███████░░░ 14/20 Adequate ── Overall Assessment ───────────────────────────────────── total ███████░░░ 73/100 🔴 High Risk (prone to hallucination) [Recommendation] Provide sufficient background, prerequisites, or few-shot examples to ground model output. ── [Eco-Conscious] Carbon Footprint Analysis ── Boundary / Boundedness ██████████ 20/20 Naturally abundant Manifestation / Clarity ████████░░ 16/20 Naturally abundant Substrate / Grounding ████░░░░░░ 9/20 Needs reinforcement Regulation / Coherence ███████░░░ 14/20 Moderate natural supply Circulation / Context Flow ███████░░░ 14/20 Moderate natural supply ── [Multi-Dimensional] Semantic & Structural ── Boundedness / 定界 ████████░░ 16/20 Strong Clarity / 顯化 ███████░░░ 14/20 Moderate Grounding / 承載 ████░░░░░░ 8/20 Weak Coherence / 定律 ███████░░░ 14/20 Moderate Context Flow / 流通 ███████░░░ 13/20 Moderate ── Green Energy Cycle Detail ── (Ecological balance: no additional energy compensation required) Raw carbon score: 86 → Post-green-optimization: 86 ── Energy Efficiency Summary ────────────────────────────── total ████████░░ 86/100 Near carbon-neutral ⏱ Performance (Timing) ├─ Keyword scan: 145.95ms ├─ Model load: 606.03ms ├─ Semantic inference: 4.02s └─ Total: 4.17s $ _

Vegan mode: Carbon footprint / eco-conscious language package · Specialized for environmental sustainability metrics

真實對話記錄 · Live Session
Live Session · 真實對話記錄

這不是示範。
這是它自然發生的樣子。

This wasn't staged.
This is what it looks like when it happens naturally.

—— 對話節錄

從一個關於「AI 業內鄙視鏈」的玩笑開始,對話自然走到了五行結構分析, 走到了「5Q 調的是人不是模型」,走到了「魔鬼不要錢,他要人墮落」, 再走到產品定位與行銷策略——中間沒有斷點,沒有人說「讓我示範一下」。

—— Conversation excerpt

Starting from a joke about "AI industry pecking orders," the conversation naturally moved into Five-Element structural analysis, then to "5Q calibrates the person, not the model," then to "the devil doesn't want money — he wants you to fall," and on to product positioning and marketing strategy — no breakpoints, no one said "let me demonstrate."

誠實說明:這段對話的深度由三件事共同造成—— 預先載入的 5Q 理論框架作為系統脈絡、使用者本身的提問品質、以及 AI 的底層能力。 這不是診斷工具的單獨效果,而是脈絡設計與提問品質同時到位時,會自然發生的事。

完整記錄開放閱讀,不刪節,不美化。讀者可以自己判斷:深度在哪一刻產生,因為什麼。

Honest disclosure: the depth of this conversation was produced by three things together — the 5Q framework preloaded as system context, the quality of the user's questions, and the AI's underlying capability. This is not the diagnostic tool alone; it's what naturally happens when context design and question quality both arrive at once.

The full record is open to read, unedited and unembellished. Readers can judge for themselves: where did depth appear, and why.

煉化資助 · Support the Research
Support the Research · 煉化資助

若想對此研究有所支持

If you'd like to support this work