Prompt Alchemy

Compress vague, verbose prompts into precise concept-driven instructions. Based on the insight that a well-chosen framework name activates more knowledge in an LLM than a thousand words of description.

Modes

Execution Model

Use a single isolated Agent for forward mode to protect main context. The agent receives: user intent summary + selected strategy + loaded cards + output contract. Main context only sees the final result (~200 tokens residual).

For reverse and deep modes, execute inline (no agent needed for lookups).

Forward Mode Flow

Step 1: Task Profile (internal, not displayed)

Assess 3 dimensions from user input:

• goal_mode: style (tone, voice, creativity) / structure (framework, organization, logic) / truth (accuracy, evidence, constraints)
• ambiguity: low / mid / high
• domain_sidecar: none / career / cv / knowledge

Step 2: Select Strategy + Constraints

Pick two things:

• strategy: The reasoning framework or method (e.g., "Pyramid Principle", "STAR method", "Porter's Five Forces")
• constraints: Audience + format + length + tone derived from context

Rules:

• structure-dominant → framework/theory keys preferred
• style-dominant → persona keys allowed as overlay only, never as primary reasoning engine
• truth-dominant → no persona keys; use "concept + explicit constraints + verification"
• When ambiguity is high → append 1 clarifying question OR flag the ambiguity in output

Step 3: Output

Default format (inline annotation):

🔑 [Refined prompt here]
   [strategy: goal_mode → chosen_framework | not chosen: alternative (1-line reason)]

Then immediately produce the result — don't wait for confirmation.

Example:

🔑 Use STAR method to rewrite this internship experience, quantify outcomes, target risk/quant roles
   [strategy: structure → STAR | not chosen: CAR (too compact, loses situational context)]

[...rewritten content follows directly...]

Step 4: Deep Mode (/alchemy:deep)

Only when explicitly requested. Show:

• Task profile (all 3 dimensions)
• 2-3 candidates with tradeoff comparison
• Hit signals and reasoning
• Full execution contract (strategy + constraints + verification)

Reverse Mode (/alchemy:rev)

Input: a concept/framework name (e.g., "费曼学习法", "Porter's Five Forces")

Output:

🔄 [Concept Name]

What it is: [2-3 sentences]

Good for:
  1. [scenario + example]
  2. [scenario + example]

Not good for:
  1. [scenario + reason]

Prompt templates:
  - [template 1]
  - [template 2]

Related: [1-2 related concepts with one-line explanation]

First check .onekey/cards/ for a matching card. If found, use card data. If not, use model knowledge.

Knowledge Base

Location: Project root .onekey/

.onekey/
  index.tsv          # concept index (name | aliases | domain | type)
  cards/             # one .md per concept, 80-180 tokens each
  local/             # cv/career sidecar indexes
    profile-core.md
    cv-index.tsv
    career-index.tsv

Loading rules:

• L1/L2 triggers: read index.tsv → grep match → load 1-3 relevant cards
• L3 (intent heuristic): do NOT load index; rely on model knowledge only
• Sidecar: load profile-core.md + relevant cv/career rows only when task is career/cv related

Index is never loaded into main context permanently. Only read inside the isolated agent.

Intent Heuristic (L3 trigger — used by CLAUDE.md routing)

When no explicit /alchemy command or keyword is present, assess input using weighted signals:

Positive signals:

• +3 Asks to "improve/polish/make more X" without specifying a method
• +2 Rewrite/polish/revise task type
• +2 Has goal/effect but missing audience/format/length/tone (≥2 missing)
• +1 2+ vague adjectives ("professional", "natural", "compelling")
• +1 Contradictory constraints ("formal but casual", "brief but thorough")
• +1 Recent 2 turns were about expression/writing optimization

Negative signals:

• -3 User already specified a method/framework
• -3 "Just do it, don't explain"
• -2 Pure translation / summary / extraction / code fix
• -2 All constraints already provided (audience + format + length + tone)
• -1 "Don't ask, just write"

Short-input special case: "rewrite verb + subjective goal + missing method/audience/format" → candidate even at 15 words.

Thresholds:

• ≥4: Inline alchemy annotation in output (don't ask, just annotate)
• 2-3: No annotation, no interruption (not confident enough)
• ≤1: Normal execution

Token Budget

What This Skill Is NOT

• Not a prompt engineering tutorial (that's the prompt-engineering skill)
• Not a writing assistant (it refines the instruction, then executes)
• Not automatic for every input (L3 only fires at score ≥4, and just annotates)