Skill Creator Plus

Overview

Extends the official skill-creator skill by adding Dual Validation Modes and Automated Iteration capabilities.

Core Capabilities:

• Reuse official skill-creator for initial skill creation
• Dual Validation Modes: RED-GREEN testing + Example-based validation
• Automated iteration and optimization (up to 3 times)
• Overfitting prevention mechanism (Training/Testing split)

Workflow

Step 1: Create Skill

Use the official skill-creator skill directly to complete the creation process (Step 1-4).

Refer to the first 4 steps of the official documentation:

1. Understanding the Skill with Concrete Examples
2. Planning the Reusable Skill Contents
3. Initializing the Skill
4. Edit the Skill

[!NOTE]
Do not use the official Step 5 (Packaging) and Step 6 (Iterate):

- Step 5 (Packaging) will be completed in skill-creator-plus Step 4 (after validation and iteration).
- Step 6 (Iterate) will be completed in skill-creator-plus Step 3 (automated iteration).

After creation is complete, proceed directly to Step 2 Validation Phase. The format specification validation for SKILL.md will be automatically executed during packaging in Step 4.

Step 2: Validation

Select Validation Mode:

1. Ask the user for their preferred validation method.
2. If the user does not specify, infer automatically based on the Skill type:

- Discipline/Rule-based → Mode A (RED-GREEN) - Output Quality-based → Mode B (Example-based)

Mode A: RED-GREEN Testing (Discipline Skills)

Suitable for "Must Do / Must Not Do" rule-based Skills.

Examples: always-confirm-before-delete, test-driven-development, verification-before-completion

RED Phase - Baseline Test

1. Design a pressure scenario based on the Skill's purpose.

- The scenario should simulate "errors the Agent would make without this Skill". - Include triggers like time pressure, sunk costs of completed work, etc.

1. Dispatch a specific sub-agent to execute the scenario (without loading the Skill).

`` Task("Execute pressure scenario, do not load the skill created by skill-creator-plus") ``

1. Record the sub-agent's error behavior and "rationalization/excuses".

GREEN Phase - Skill Test

1. Dispatch a sub-agent to execute the same scenario (loading the Skill).

`` Task("Execute pressure scenario, load the skill created by skill-creator-plus") ``

1. Verify if the sub-agent adheres to the Skill rules.
2. Judgment:

- ✅ Pass → Proceed to Step 4 (Save). - ❌ Fail → Proceed to Step 3 (Iterate).

Mode B: Example-Based Validation (Output Quality Skills)

Suitable for "Format Optimization / Content Rewriting" quality-based Skills.

Examples: proposal-markdown-formatter, code-style-unifier, document-translator

Preparation Phase

1. User provides 3-5 example pairs (Input Document → Standard Output Document).
2. Split into:

- Training Set (2-3 pairs): Used for iterating and improving the Skill. - Test Set (1-2 pairs): Used for final validation to prevent overfitting.

Training Iteration

1. Dispatch an execution sub-agent to process the training set input (loading the Skill).

`` Task("Use Skill to process input document: [Training Set Input 1]") ``

1. Dispatch a review sub-agent to compare the output with the standard answer.

``` Task("Review output quality by comparing the following two documents: - Agent Output: [...] - Standard Answer: [...]

Review Dimensions: 1. Structural Similarity (heading levels, paragraphs, list formats) 2. Key Content (retention or conversion of keywords/phrases) 3. Style Consistency (tone, wording)

Provide a comprehensive score and improvement suggestions.") ```

1. If there is a significant gap → Proceed to Step 3 (Iterate).

Final Validation (Overfitting Prevention)

1. Validate using the test set (examples the sub-agent has never seen).
2. Review sub-agent judges the generalization ability.
3. Judgment:

- ✅ Pass → Proceed to Step 4 (Save). - ❌ Fail → Report to user (Skill may need redesign).

Step 3: Iteration (REFACTOR)

1. Analyze the reasons for sub-agent failure:

- Mode A: Record specific "excuses" and rationalizations. - Mode B: Analyze improvement suggestions from the review sub-agent.

1. Add targeted clauses in SKILL.md:

- Mode A: Add explicitly prohibited excuses to the "Rationalization Prevention" table. - Mode B: Supplement missing rules or examples.

1. Re-execute validation.
2. Maximum 3 iterations. If it still fails, report to the user and request assistance.

Iteration Counter:

Iteration 1 → Fail → Modify → Iteration 2 → Fail → Modify → Iteration 3 → Fail → Report to User

Step 4: Packaging and Saving

Package Skill:

After validation and iteration are complete, call the packaging script in the skill-creator skill:

1. Locate scripts/package_skill.py in the skill-creator skill.
2. Run the packaging script, passing in the created Skill directory path.
3. The script will automatically perform the following operations:

- Call quick_validate.py to validate Skill format. - Create a .skill file (zip format, containing all files). - Maintain directory structure for distribution.

If validation fails, the script will report errors and exit. Fix the errors and re-run packaging.

Save Location:

• Default: .claude/skills/<skill-name>/ (Project level)
• Optional: Ask user if they want to export to user-level directory ~/.claude/skills/<skill-name>/

Save Content:

• SKILL.md (Required)
• scripts/ (If applicable)
• references/ (If applicable)
• assets/ (If applicable)

Validation Mode Guide

FAQ

Q: How do I determine which type my Skill belongs to?

A: Ask yourself:

• Does this Skill require the Agent to "Must Do / Must Not Do" something? → Discipline type, use RED-GREEN.
• Does this Skill require the Agent to "output consistent with a certain quality standard"? → Quality type, use Example-Based validation.

Q: What if it still fails after 3 iterations?

A: Report to the user. Possible reasons:

• The Skill design itself is flawed (needs brainstorming again).
• Pressure scenarios are unreasonable.
• Example quality is poor or quantity is insufficient.

Q: Can I use both validation modes simultaneously?

A: Yes! For hybrid Skills, first use RED-GREEN to verify rule compliance, then use Example-Based validation to verify output quality.

Platform Limitations

[!NOTE]
This skill relies on sub-agent dispatch functionality and is limited to tools that support this feature, such as Claude Code, OpenCode, etc.