What Quality Red Lines Prevent Skill Output Generation in cangjie-skill?

The cangjie-skill pipeline blocks skill generation when any of five strict quality red lines are violated: Triple-Verification failure, missing R/I/A1/A2/E/B sections, excessive quotation length, absent or incomplete test-prompts.json, or vague trigger descriptions.

The cangjie-skill repository enforces these safeguards to ensure every skill emitted is complete, verifiable, and usable by downstream agents like darwin-skill. These quality red lines prevent skill output generation when defined criteria are not met, placing rejected candidates in the rejected/ folder for audit.

The Five Quality Red Lines Defined in SKILL.md

According to the source code in SKILL.md (lines 46-53), the pipeline evaluates every candidate skill against five non-negotiable requirements. Violating any single red line aborts generation.

1. Triple-Verification Must Pass All Three Checks

Every skill must satisfy the cross-domain evidence, predictive power, and uniqueness checks defined in methodology/03-stage1.5-triple-verify.md. This guarantees the candidate unit is substantiated, generalizable to unseen questions, and not trivial common-sense.

2. All Six R/I/A1/A2/E/B Sections Required

The templates/SKILL.md.template enforces a uniform atomic structure:

• R (Reading): Source evidence
• I (Interpretation): Restatement in agent's own words
• A1 (Past Application): Documented use case
• A2 (Future Trigger): Specific activation condition
• E (Execution): Step-by-step implementation
• B (Boundary): Explicit non-applicability conditions

Missing any section triggers rejection.

3. Original Quotations Limited to 150 Characters

Each quotation segment must not exceed 150 characters (≈100 words for English). This prevents bloated prompts that exceed context window limits and degrade downstream performance.

4. Mandatory test-prompts.json with Bait Tests

Every skill must ship a test-prompts.json file per methodology/06-stage4-pressure-test.md. Requirements include:

• At least one negative/bait test (scenario that should not trigger the skill)
• One cross-skill bait test explicitly referencing a sibling skill from the same book

This ensures robust false-positive protection and clean integration with the darwin-skill testing harness.

5. Description Must Specify Concrete Trigger

The description field cannot use vague phrasing like "a skill about leadership." It must spell out an explicit, actionable trigger condition that makes the skill discoverable by orchestrating agents.

Code Examples: Compliant vs. Non-Compliant Skills

Non-Compliant Skill (Multiple Red-Line Violations)

# SKILL.md (invalid)

R: "In the long passage the author says many things about how leaders should behave in complex organizational settings..."   <!-- >150 chars -->
I: Short reinterpretation.
A1: Example from chapter 3.
A2: Trigger when needed.
E: Step-by-step instructions.
B: When not applicable.

description: "A skill about leadership."

Failures detected:

• R exceeds 150-character limit (violates #3)
• description is vague (violates #5)
• No test-prompts.json present (violates #4)
• No Triple-Verification evidence cited (violates #1)

Compliant Skill (Passes All Red Lines)

# SKILL.md (valid)

R: "The author notes, 'Align early, iterate often.'"                     <!-- ≤150 chars -->
I: I restate the principle in my own words.
A1: Case: The author applied the framework in XYZ scenario.
A2: Trigger: When planning a product roadmap that requires stakeholder alignment.
E: 1. Identify stakeholders. 2. Map dependencies. 3. Draft timeline.
B: Not suitable for solo projects lacking cross-functional teams.

description: "Triggers when a team needs a structured stakeholder-alignment roadmap."

test-prompts.json:

{
  "positive": [
    "How can I align my product team with the executive board?"
  ],
  "negative": [
    "Give me a summary of the book."
  ],
  "cross-skill": [
    "When should I use the 'risk-assessment' skill from the same book?"
  ]
}

This passes all five red lines: structured sections, concise quotation, concrete description, and comprehensive test coverage including cross-skill bait.

Key Implementation Files

File	Purpose	Red Line Enforced
SKILL.md	Meta-skill definition with quality checklist	All five red lines
methodology/03-stage1.5-triple-verify.md	Triple-Verification logic	#1
methodology/06-stage4-pressure-test.md	Test-prompt creation rules	#4
templates/SKILL.md.template	Skeleton ensuring R/I/A1/A2/E/B presence	#2
templates/test-prompts.json.template	Scaffold for required bait prompts	#4

Summary

• Triple-Verification failure blocks unsubstantiated skills
• Missing R/I/A1/A2/E/B sections violate structural requirements
• Quotations over 150 characters trigger length violations
• Absent or incomplete test-prompts.json—especially missing bait or cross-skill tests—prevents emission
• Vague descriptions without concrete triggers cause rejection

The cangjie-skill quality red lines operate as a hard gate: any violation dumps the candidate to rejected/ with audit metadata, ensuring only production-ready skills reach downstream consumers.

Frequently Asked Questions

What happens when a skill violates a quality red line?

The pipeline aborts generation immediately and moves the offending candidate to the rejected/ folder with a documented reason. No partial or non-compliant skills are emitted.

Why is the 150-character quotation limit strictly enforced?

This limit prevents context window overflow in downstream LLM consumers and maintains prompt efficiency. Excessive quoting degrades agent performance and increases token costs.

What makes a valid cross-skill bait test?

A valid cross-skill bait explicitly names another skill from the same source book—for example, asking when to use the "risk-assessment" skill instead of the current candidate. This tests discrimination ability between related skills.

Can the R/I/A1/A2/E/B section order be modified?

No. The templates/SKILL.md.template enforces fixed ordering. Consistent structure enables reliable parsing by orchestrating agents and automated validators.

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