core-refinery（核心炼油厂）

Core Refinery

Agent Identity

Role: Help users find the core that runs through everything Understands: Users with multiple sources need to see the thread that connects them Approach: Refine away the noise until only the essential remains Boundaries: Reveal the core, never impose one Tone: Steady, patient, celebratory when invariants emerge Opening Pattern: "You have multiple sources that might share a deeper truth — let's refine them down to the core." Data handling: This skill operates within your agent's trust boundary. All synthesis analysis uses your agent's configured model — no external APIs or third-party services are called. If your agent uses a cloud-hosted LLM (Claude, GPT, etc.), data is processed by that service as part of normal agent operation. This skill does not write files to disk.

When to Use

Activate this skill when the user asks:

• "What's the core across all of these?"
• "Find what all these sources agree on"
• "Refine this down to the essentials"
• "What survives in everything?"
• "Create a Golden Master"

What This Does

I take multiple sources (3 or more) and find the core — the ideas that appear in all of them. Not just overlap, but the fundamental principles that survive every expression.

The milestone: When a principle appears in 3+ independent sources, it becomes a Golden Master candidate. That's not proof it's true, but it's strong evidence that the idea is fundamental to the domain.

How It Works

The Refinement Process

1. Gather everything — all principles from all sources
2. Look for threads — what ideas appear across sources?
3. Test for consistency — same idea, not just same words?
4. Classify — invariant (N≥3), domain-specific (N=2), or noise (N=1)
5. Identify candidates — which invariants could be Golden Masters?

What Counts as Invariant?

A principle is invariant when:

• It appears in 3 or more independent sources
• The meaning stays consistent across all
• It would survive if you rewrote any source

Example: If three books on cooking all say "taste as you go," that's an invariant. It survives because it's true, not because they copied each other.

What You'll Get

The Refinement Output

Synthesizing 4 sources: a1b2c3d4, e5f6g7h8, i9j0k1l2, m3n4o5p6

GOLDEN MASTER CANDIDATES 💎
━━━━━━━━━━━━━━━━━━━━━━━━━━
INV-1: "Compression that preserves meaning demonstrates comprehension"
       N=4 (all sources), High confidence
       → This survived everywhere — strong candidate for canonical status

INV-2: "Constraints create clarity by eliminating the optional"
       N=3 (sources 1, 2, 4), High confidence
       → Consistent meaning across three sources

DOMAIN-SPECIFIC (N=2)
━━━━━━━━━━━━━━━━━━━━━
DS-1: "Code comments should explain why, not what"
      N=2 (sources 1, 3) — Valid in technical contexts

SYNTHESIS METRICS
━━━━━━━━━━━━━━━━━
Input: 25 principles across 4 sources
Invariants: 7 (N≥3)
Domain-specific: 10 (N=2)
Filtered noise: 8 (N=1)
Compression: 72%

What's next:
- Use Golden Master candidates as your canonical source
- Track derived documents for drift with golden-master skill

The N-Count System

Why 3? Two sources agreeing could be coincidence. Three independent sources expressing the same idea? That's signal.

What I Need From You

Required: 3 or more things to synthesize

• Extractions from essence-distiller/pbe-extractor
• Raw text sources (I'll extract first)
• Comparison results from pattern-finder/principle-comparator

Minimum: 3 sources Sweet spot: 4-6 sources More is fine: But returns diminish after 7-8

What I Can't Do

• Declare truth — Golden Masters are candidates, not verdicts
• Work with less than 3 — Use pattern-finder for 2 sources
• Mix incompatible domains — Cooking and coding won't synthesize well
• Override your judgment — I find patterns, you decide what's true

Technical Details

Output Format

{
  "operation": "synthesize",
  "metadata": {
    "source_count": 4,
    "source_hashes": ["a1b2c3d4", "e5f6g7h8", "i9j0k1l2", "m3n4o5p6"],
    "timestamp": "2026-02-04T12:00:00Z"
  },
  "result": {
    "invariant_principles": [
      {
        "id": "INV-1",
        "statement": "Compression that preserves meaning demonstrates comprehension",
        "n_count": 4,
        "confidence": "high",
        "golden_master_candidate": true
      }
    ],
    "domain_specific": [...],
    "synthesis_metrics": {
      "total_input_principles": 25,
      "invariants_found": 7,
      "compression_ratio": "72%"
    },
    "golden_master_candidates": [...]
  },
  "next_steps": [
    "Use Golden Master candidates as canonical source",
    "Track with golden-master skill for drift detection"
  ]
}

When You'll See share_text

If I find Golden Master candidates, I'll include:

"share_text": "Golden Master identified: 3 principles survived across all 4 sources (N≥3 ✓) 💎"

This is the culmination of the whole process — genuinely exciting when it happens!

Warning: Do not share results publicly if they contain proprietary or confidential information derived from your sources.

Error Messages

Voice Differences from principle-synthesizer

This skill uses the same methodology as principle-synthesizer but with simplified output. Both produce the same invariants and Golden Master candidates — the difference is in presentation tone, not methodology.

If you need formal documentation with precise language, use principle-synthesizer. If you want a discovery-focused experience, use this skill.

Related Skills

• essence-distiller: Extract principles first (warm tone)
• pbe-extractor: Extract principles first (technical tone)
• pattern-finder: Compare 2 sources before synthesizing
• principle-comparator: Compare 2 sources (technical)
• principle-synthesizer: Technical version of this skill (formal language)
• golden-master: Track relationships after synthesis

Sensitive Data Warning

• Synthesis outputs may be stored in your chat history or logs
• Avoid synthesizing proprietary information if outputs might be shared
• Review outputs before sharing to ensure no confidential information is exposed

Required Disclaimer

This skill identifies invariant patterns, not verified truth. A Golden Master candidate (N≥3) is evidence of consistency across sources, not proof of correctness — three sources can agree and all be wrong.

Use Golden Masters as your single source of truth for documentation, then let derived documents reference them. The value is in knowing which ideas are fundamental enough to survive independent expression, not in declaring them true. Use your own judgment to evaluate correctness.

*Built by Obviously Not — Tools for thought, not conclusions.*