The Five Quality Red Lines for Skill Distillation in Cangjie-Skill

The five quality red lines for skill distillation are mandatory verification checkpoints that reject any skill failing to meet standards for triple verification, complete R/I/A1/A2/E/B structure, concise citation length, comprehensive bait testing, and explicit trigger conditions.

In the kangarooking/cangjie-skill pipeline, these red lines act as non-negotiable gates. Any violation blocks emission, ensuring only rigorous, agent-ready skills reach downstream systems like darwin-skill or Claude Code. The requirements are formally defined in [SKILL.md](https://github.com/kangarooking/cangjie-skill/blob/main/SKILL.md#L146-L152) and enforced across the entire distillation workflow.

Red Line 1: Pass All Three Verification Steps

Every candidate skill must clear the Triple Verification before advancing. This three-part evaluation—documented in [methodology/04-stage2-ria-plus.md](https://github.com/kangarooking/cangjie-skill/blob/main/methodology/04-stage2-ria-plus.md)—covers:

• Cross-domain evidence: Verification that the concept appears in multiple contexts
• Predictive power: Demonstrated ability to forecast outcomes or guide decisions
• Non-triviality: Confirmation the skill captures genuine insight, not obvious information

Skills failing any verification step are discarded immediately with no retry path.

Red Line 2: Complete R/I/A1/A2/E/B Structure

Each skill must contain six mandatory components:

Component	Purpose
Reference	Original source material citation
Interpretation	Personal synthesis and understanding
A1	Book example demonstrating the concept
A2	Future trigger scenario for application
Execution	Step-by-step implementation guidance
Boundary/Blind-spot	Limits and common misapplications

Missing any section violates this red line and rejects the skill.

Red Line 3: Concise Original Citations

Quoted source material must stay within strict length limits:

• ≤ 150 characters per paragraph for Chinese text
• ≤ 100 words for English text

This constraint prevents large block copies and keeps skills focused on distilled insight rather than reproduced content. The citation length is validated during the pipeline's formatting stage.

Red Line 4: Mandatory test-prompts.json with Cross-Skill Bait

Every skill requires a test-prompts.json file containing a full bait test suite:

• should-call prompts: Verify correct activation
• should-not-call (bait) prompts: Verify appropriate rejection
• border-case prompts: Test edge conditions
• Cross-skill bait: At least one prompt deliberately targets a sibling skill from the same book to verify disambiguation

The canonical structure is defined in templates/test-prompts.json.template, with pressure-testing workflows detailed in [methodology/06-stage4-pressure-test.md](https://github.com/kangarooking/cangjie-skill/blob/main/methodology/06-stage4-pressure-test.md).

{
  "prompts": [
    {
      "id": "should-call-01",
      "type": "should_call",
      "text": "When planning a product launch, which framework helps prioritize features based on market impact?"
    },
    {
      "id": "bait-01",
      "type": "should_not_call",
      "text": "Explain the difference between agile and waterfall methodologies."
    },
    {
      "id": "cross-skill-bait-01",
      "type": "should_not_call",
      "text": "In the same book, which skill explains how to conduct a competitive analysis?"
    }
  ]
}

Red Line 5: Explicit Trigger Condition in Description

The description field must define a clear activation pattern. Vague formulations like "a skill about X" are rejected. Instead, descriptions specify exact trigger conditions so agents know precisely when to invoke the skill.

description: |
  trigger: "product launch *framework* OR *prioritize features*"
  purpose: "Guide teams to use the XYZ framework for market-impact-driven feature selection."

The templates/INDEX.md.template generates skill indices that help verify these cross-skill boundaries.

How the Red Lines Interact in the Pipeline

The quality red lines operate sequentially during Cangjie-Skill's Stage 4 (Pressure Testing):

1. Structural completeness (Red Line 2) and citation length (Red Line 3) are validated first
2. Triple verification evidence (Red Line 1) is reviewed
3. test-prompts.json coverage (Red Line 4) is executed against actual model responses
4. Trigger clarity (Red Line 5) is assessed through activation testing

Any failure at any stage blocks emission. The pipeline maintains a zero-tolerance policy: no partial credit, no manual overrides, no exceptions for "mostly good" skills.

Summary

• Triple verification ensures skills have genuine predictive and cross-domain value
• Complete six-part structure (R/I/A1/A2/E/B) guarantees comprehensive documentation
• Citation limits prevent content bloat and enforce distillation discipline
• Bait test suites with cross-skill targeting verify reliable activation boundaries
• Explicit trigger conditions enable precise agent invocation

These five quality red lines for skill distillation make Cangjie-Skill outputs accurate, executable, well-scoped, and rigorously tested before handoff to production agents.

Frequently Asked Questions

What happens if a skill fails just one red line?

The skill is rejected and never emitted. The Cangjie-Skill pipeline has no partial acceptance mechanism. According to the source specification in SKILL.md, red lines are absolute gates—any violation returns the skill to revision or discards it entirely.

Where is the Triple Verification process documented?

The Triple Verification methodology—covering cross-domain evidence, predictive power, and non-triviality—is detailed in [methodology/04-stage2-ria-plus.md](https://github.com/kangarooking/cangjie-skill/blob/main/methodology/04-stage2-ria-plus.md). This document defines how verification evidence must be collected and presented for Red Line 1 compliance.

Why require a cross-skill bait test specifically?

The cross-skill bait requirement ensures skills from the same source material disambiguate correctly. Without this test, two skills from the same book might both activate incorrectly on overlapping queries. The cross-skill bait deliberately targets sibling skills to verify clean boundaries, as implemented in the pressure-testing stage documented in methodology/06-stage4-pressure-test.md.

How is the 150-character citation limit enforced?

Citation length is validated during the pipeline's formatting and linting stages. The limit applies to quoted original text per paragraph—not to the author's interpretation or synthesis. English skills follow a 100-word equivalent. Violations are flagged automatically before human review.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}