How Boundary Conditions (B Field) Prevent Skill Misuse in Cangjie-Skill

The B (Boundary) field in cangjie-skill explicitly defines when a skill should not be applied, filtering out inappropriate contexts before execution to prevent "random calls" and maintain reliability.

The cangjie-skill repository implements a structured methodology for distilling knowledge into reusable AI skills. Each skill follows the R-I-A1-A2-E-B format, where the B (Boundary) segment serves as a critical guardrail against misuse. By documenting explicit non-applicability conditions and author blind-spots, the boundary conditions ensure that skills are only invoked in appropriate contexts, eliminating the risk of over-generalized or erroneous outputs.

Understanding the B (Boundary) Field Structure

The cangjie-skill framework organizes every distilled skill into six distinct sections: R (Role), I (Input), A1 (Analysis 1), A2 (Analysis 2), E (Execution), and B (Boundary). This structure ensures comprehensive coverage of both applicability and limitations.

The Six-Section Skill Architecture

In methodology/04-stage2-ria-plus.md, the framework defines the B segment as the repository for when the skill should NOT be applied and any blind-spots identified during the original source's critique phase. Without this section, a skill lacks the necessary constraints to prevent inappropriate invocation.

The methodology explicitly warns that missing boundary conditions lead to skill overuse: "没边界的 skill 会被过度调用,最终用户失望" (skills without boundaries will be over-called, ultimately disappointing users) 【methodology/04-stage2-ria-plus.md#L89】.

How Boundaries Are Derived from Source Analysis

Boundary conditions do not emerge arbitrarily; they are systematically extracted from the source material's critique phase and counter-example analysis.

Extracting Counter-Examples

According to extractors/counter-example-extractor.md, counter-examples are essential for establishing valid boundaries. The documentation states: "没有反例,skill 就没有边界,会在不该用的时候被调用,反而帮倒忙" (without counter-examples, skills have no boundaries and will be called when they shouldn't be, causing more harm than good) 【extractors/counter-example-extractor.md#L7】.

These counter-examples feed directly into the B field, creating explicit exclusion criteria based on real limitations discovered during the distillation process.

The BOOK_OVERVIEW Critique Phase

The initial critique phase (Stage 0) generates the BOOK_OVERVIEW, which identifies author blind-spots and contextual limitations. These findings are preserved in the B section to prevent the AI from generalizing beyond the source material's valid domain.

Preventing Skill Misuse at Runtime

The B field operates as an active filter during skill selection, not merely passive documentation.

The Applicability Check Logic

When an AI agent receives a request, it evaluates candidate skills against their Boundary conditions before execution. The following pseudo-logic illustrates this guardrail mechanism:

def is_skill_applicable(skill, request):
    # Iterate through B field conditions

    for condition in skill.boundary.not_applicable:
        if condition.matches(request):
            return False  # B field blocks misuse

    return True

This check occurs immediately before the E (Execution) steps run. If the request matches any "not-applicable" condition listed in the B section, the skill is filtered out, ensuring only context-appropriate knowledge surfaces.

Explicit Non-Applicability Documentation

A typical B section in SKILL.md explicitly lists exclusion scenarios:

## B (Boundary)

- **不适用场景**：当用户询问关于"实时股价"而不是"投资原则"时，本 skill 不应被调用。
- **盲点**：原作者在书中未讨论新兴的加密资产，故本 skill 不覆盖此类资产的评估。

This concrete definition prevents the "random calls" (防止乱调用) highlighted in the methodology 【methodology/04-stage2-ria-plus.md#L68】.

Enforcing Boundaries Through Templates

The templates/SKILL.md.template and the generated SKILL.md files enforce boundary documentation through structural requirements. Line 119 of the standard skill template includes a dedicated B (Boundary) bullet prompting authors to list "什么时候不适用 / 来自阶段 0 批判阶段的作者盲点" (when not to apply / author blind-spots from stage 0 critique) 【SKILL.md#L119】.

This template-driven approach ensures that no skill enters the repository without explicit boundary definitions, institutionalizing the guardrail against misuse.

Summary

• The B (Boundary) field in cangjie-skill explicitly records when skills should not be applied and identifies author blind-spots from the critique phase.
• Boundary conditions are derived from counter-examples and the BOOK_OVERVIEW critique, ensuring they reflect actual limitations rather than arbitrary restrictions.
• The methodology warns that skills without boundaries risk being over-called, leading to user disappointment and erroneous outputs.
• Runtime applicability checks filter skills based on B field conditions before executing E (Execution) steps, preventing inappropriate invocations.
• The SKILL.md template enforces boundary documentation, ensuring every skill includes explicit non-applicability scenarios.

Frequently Asked Questions

What happens if a skill lacks a B (Boundary) section?

According to the cangjie-skill methodology in methodology/04-stage2-ria-plus.md, skills without boundaries will be "over-called" (过度调用), resulting in disappointed users. Without explicit non-applicability conditions, the AI agent cannot filter out inappropriate contexts, leading to random invocations that apply knowledge outside its valid domain.

How are boundary conditions different from input validation?

While input validation checks format and syntax, the B field evaluates semantic and contextual applicability. It answers whether the skill's knowledge applies to the user's specific scenario based on blind-spots and limitations identified during source analysis, not merely whether the input structure is correct.

Where does the content for the B field originate?

Boundary content comes from two primary sources documented in the repository: counter-examples collected during Stage 2 (as noted in extractors/counter-example-extractor.md) and the BOOK_OVERVIEW critique from Stage 0. These sources identify when the original author's advice does not apply or where their knowledge has gaps.

Can boundary conditions be updated after initial skill creation?

Yes, the methodology implies that as new counter-examples emerge or additional blind-spots are identified, the B section in SKILL.md should be updated to reflect these limitations. This iterative refinement ensures the skill guardrails remain accurate as contexts evolve.

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