How to Create Effective Execution Steps (E Field) for Agents in Cangjie-Skill

The E field converts abstract methodology into a concrete 1-2-3 step workflow with clear completion criteria, enabling deterministic, auditable agent execution.

The cangjie-skill repository defines a structured framework for building agent-ready skills, and the E field (Execution) serves as the deterministic core that dictates exactly what actions an agent must perform. Unlike free-form instructions, a properly crafted E field provides a sequential, testable workflow that agents can execute step-by-step without ambiguity. This guide explains how to construct effective execution steps based on the official templates and methodology defined in the kangarooking/cangjie-skill codebase.

What Is the E Field?

The E field (可执行步骤) is the execution layer of a cangjie-skill that translates abstract methodology into concrete agent actions. According to templates/SKILL.md.template, this section must contain a numbered list of steps that the agent follows sequentially once the skill is triggered.

The primary purpose is to eliminate non-deterministic behavior by providing:

• Clear completion criteria (完成标准) for each step
• Optional stop conditions (判停条件) that redirect flow when specific criteria are met
• Deterministic branching that prevents endless loops or false activations

This structure ensures that sub-agents can execute the skill during the pressure-test stage (documented in methodology/06-stage4-pressure-test.md) without requiring external hints or interpretation.

Structure Requirements from the Official Template

The canonical structure is defined in templates/SKILL.md.template. Every E field must follow this exact markdown pattern:

## E — 可执行步骤 (Execution)

当 skill 被激活后, agent 应按以下步骤执行:

1. **{{步骤 1}}**
   - 完成标准: {{如何判断这一步已完成}}

2. **{{步骤 2}}**
   - 完成标准: {{ ... }}
   - 判停条件: 若 {{X}} 则跳到步骤 {{N}}

Key structural requirements include:

• Bold step titles using double asterisks
• Hyphenated sub-items for completion standards and stop conditions
• Sequential numbering (1, 2, 3) to enforce order
• Explicit jump syntax for stop conditions ("若 X 则跳到步骤 Y")

Core Design Principles from the Methodology

The design rules in methodology/04-stage2-ria-plus.md codify three essential principles for writing execution steps:

"把方法论转成 1-2-3 步骤" — Convert methodology into short, sequential steps. Avoid convoluted branching; keep the flow linear where possible.

"每一步有可判断的完成标准" — Every step must have a measurable completion condition that the agent can verify independently. Vague criteria like "保持客观" (remain objective) violate this rule.

"如果有判停点 … 显式写出" — Embed branching logic directly within the step description using explicit stop conditions rather than external logic.

These principles ensure the skill is executable rather than a theoretical discussion, enabling automated testing and integration with darwin-skill.

Writing Execution Steps: A Practical Guide

Define Concrete Actions

Start by identifying 1-3 concrete actions that transform input into output. According to the template, each step title should describe a specific verb-object action.

1. **列出用户的核心目标**
   - 完成标准: 成功提取并输出至少 3 条用户在提示中暗示的目标。

This example from the repository demonstrates a measurable extraction task with a quantifiable success condition (≥3 goals).

Specify Measurable Completion Criteria

The 完成标准 (completion standard) must be objective and verifiable. The agent should determine success without human intervention.

Good example:

• 完成标准: 从用户提供的背景信息中抽取 ≥4 条时间点。

Bad example:

• 完成标准: 理解用户的需求。

Add Stop Conditions for Branching

When a step requires conditional logic, use the 判停条件 (stop condition) syntax to redirect execution flow explicitly.

2. **识别最不希望发生的结果**
   - 完成标准: 输出 1-2 条"最不希望的后果"，并在用户确认后继续。
   - 判停条件: 若用户回答 "没有最不想要的"，则跳到步骤 3。

This pattern prevents the agent from getting stuck in infinite loops or executing unnecessary steps when preconditions aren't met.

Integration with Pressure Testing

The precision of your E field directly impacts the pressure-test stage described in methodology/06-stage4-pressure-test.md. During this phase, a sub-agent receives only:

• The skill description
• The user prompt
• Optional neighboring skills list

The sub-agent must decide whether to trigger the skill and how to act. A well-crafted E field supplies the exact actions the sub-agent can perform, reducing false activations and ensuring deterministic behavior in production environments.

Common Pitfalls to Avoid

Avoid these errors that invalidate execution steps:

• Missing completion criteria — Steps without 完成标准 leave the agent unable to determine when to proceed.
• Vague action descriptions — Abstract concepts without measurable outcomes violate the "every step has a verifiable completion standard" rule.
• Implicit branching — Hidden conditional logic that isn't declared as a 判停条件 leads to unpredictable execution paths.
• Excessive step count — While more than 3 steps are allowed, lengthy workflows reduce determinism and complicate pressure testing.

Summary

• The E field in kangarooking/cangjie-skill converts abstract methodology into a deterministic, numbered workflow for agents.
• Follow the exact structure from templates/SKILL.md.template: bold step titles, hyphenated completion criteria, and explicit stop conditions.
• Apply the three rules from methodology/04-stage2-ria-plus.md: sequential 1-2-3 steps, measurable completion standards, and explicit stop-condition syntax.
• Ensure every step includes a 完成标准 so agents can verify success independently.
• Use 判停条件 with clear "if X then jump to step Y" syntax for conditional branching.
• Validate your E field against the pressure-testing requirements in methodology/06-stage4-pressure-test.md to ensure production readiness.

Frequently Asked Questions

What makes an E field different from regular instructions?

The E field is deterministic and executable. While regular instructions might suggest an approach, the E field in cangjie-skill provides a strict 1-2-3 sequence with 完成标准 (completion criteria) that agents can verify algorithmically. According to methodology/04-stage2-ria-plus.md, this converts the skill from a discussion into an auditable workflow.

How many steps should an E field contain?

The methodology recommends 1-3 concrete steps for optimal clarity, though more steps are permitted if necessary. The key constraint is conciseness—each step must represent a discrete, measurable action with clear completion criteria. Long workflows should be decomposed into separate skills to maintain testability during the pressure-test stage.

What happens if I don't include completion standards?

Without 完成标准, the agent cannot determine when a step succeeds or fails. This violates the core principle from methodology/04-stage2-ria-plus.md that "every step has a verifiable completion standard." The agent may hang indefinitely, skip steps prematurely, or fail the pressure-testing phase where sub-agents must execute the skill autonomously.

Can I use conditional logic in the E field?

Yes, but you must use 判停条件 (stop conditions) with explicit syntax. The format "若 X 则跳到步骤 Y" (if X then jump to step Y) must be written directly under the relevant step. This ensures the branching logic is visible during pressure testing and prevents hidden execution paths that could cause infinite loops or incorrect outcomes.

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