Skills

A personal repository for collecting, organizing, and reusing Agent Skills.

This repository stores reusable skills for AI agents, coding agents, document generation agents, code review agents, tool-using agents, engineering workflows, context management, and token optimization.

Each skill is organized as an independent directory with clear usage scenarios, activation conditions, execution procedures, constraints, output requirements, and examples.

Goals

This repository is not intended to be a simple prompt collection.

Its purpose is to turn reusable agent workflows into structured, maintainable, and verifiable skills.

The main goals are:

• Capture reusable agent workflows
• Standardize skill structure and authoring conventions
• Improve the reliability of coding agents and general-purpose agents
• Reduce repetitive prompt writing
• Lower unnecessary context token usage
• Encourage agents to reuse existing tools, frameworks, and project capabilities
• Give complex tasks explicit procedures, boundaries, and failure strategies
• Provide reusable capability modules for custom agents, MCP-based agents, and Agent OS designs

What Is a Skill?

In this repository, a skill is a structured capability definition for a specific class of tasks.

A skill usually defines:

• Applicable scenarios
• Activation conditions
• Execution procedure
• Input requirements
• Output format
• Constraints
• Tool usage rules
• Failure strategy
• Examples
• Quality checklist

A skill is not the same as a one-off prompt. A prompt is usually written for a single interaction, while a skill is designed for long-term reuse, engineering consistency, and stable execution.

Repository Structure

Recommended structure:

skills/
├── README.md
├── README-ZH.md
├── framework-first-coding/
│   ├── SKILL.md
│   ├── README.md
│   └── README-ZH.md
├── agent-harness-evaluation-skill/
│   ├── SKILL.md
│   ├── README.md
│   └── README-ZH.md
└── references/
    ├── skill-template.md
    ├── writing-guide.md
    └── quality-checklist.md

Where:

• README.md is the repository-level documentation
• */SKILL.md is the core rule file used by the agent
• */README.md is the human-readable explanation for developers and maintainers
• references/ contains shared templates, authoring guides, checklists, and reference material

Included Skills

Framework-first Coding Skill

Directory:

framework-first-coding/

Purpose:

Constrains a coding agent to inspect and reuse existing framework features, dependencies, SDKs, utilities, shared components, and project abstractions before writing custom implementation code.

Suitable for:

• Adding APIs
• Fixing bugs
• Refactoring code
• Integrating third-party APIs
• Writing database queries
• Request validation
• Authorization handling
• File upload
• HTTP calls
• Logging, caching, scheduling, and other common engineering tasks

Core principle:

Find existing capabilities first. Write new code only when necessary.

This skill helps reduce duplicated implementation, lower token consumption, improve code consistency, and increase maintainability.

Agent Harness Evaluation Skill

Directory:

agent-harness-evaluation-skill/

Purpose:

Systematically evaluates the robustness, reliability, safety, recoverability, context engineering efficiency, and production readiness of AI Agent Harness implementations through deterministic testing, fault injection, chaos scenarios, and reliability metrics.

Suitable for:

• Building a new Agent framework
• Reviewing an existing Agent system before release
• Introducing new tools into the harness
• Changing retry or checkpoint mechanisms
• Upgrading model providers
• Adding multi-agent workflows
• Measuring prompt cache and context efficiency
• Preparing for production deployment

Core principle:

Evaluate the execution framework, not the model intelligence.

This skill helps detect reliability gaps before production, prevent duplicate side effects and security violations, validate recovery mechanisms, and enforce production gate criteria.

Skill Authoring Convention

Each skill should normally contain at least two files:

skill-name/
├── SKILL.md
└── README.md

SKILL.md

SKILL.md is the core file read and executed by the agent. It should be clear, actionable, and concise.

Recommended structure:

---
name: skill-name
metadata:
  version: "1.0"
  category: "category-name"
description: "One-sentence description of when and why to use this skill."
---

# Skill Name

## Purpose

## When to Use

## Inputs

## Execution Procedure

## Output Requirements

## Constraints

## Tool Usage Rules

## Failure Strategy

## Quality Checklist

## Examples

README.md

README.md is written for humans. It can explain the background, purpose, suitable scenarios, behavior, and examples in more detail.

Recommended structure:

# Skill Name

## Overview

## Why This Skill Exists

## Suitable Use Cases

## How It Works

## Expected Agent Behavior

## Directory Structure

## Usage

## Examples

## Limitations

## Recommended Pairing

Naming Convention

Skill directory names should use lowercase English words and hyphens:

framework-first-coding
minimal-patch-coding
code-review
security-review
document-generation
tool-failure-strategy

Avoid:

FrameworkFirstCoding
framework_first_coding
框架优先编码
skill1
my-skill-new

Naming principles:

• Short
• Clear
• Searchable
• Describes the core capability
• Avoids unnecessary abstraction

Suggested Skill Categories

Skills can be grouped by task type.

Coding Skills

For code generation, modification, refactoring, and review.

Examples:

• framework-first-coding
• minimal-patch-coding
• code-review
• security-review
• test-first-coding
• dependency-inspection
• project-convention-following

Agent Skills

For agent execution flow, tool use, and task orchestration.

Examples:

• tool-use-planning
• tool-failure-strategy
• multi-step-task-execution
• human-in-the-loop
• agent-loop-control
• context-compaction

Document Skills

For creating, editing, summarizing, and structuring documents.

Examples:

• document-generation
• technical-report-writing
• research-summary
• docx-processing
• slide-generation
• proposal-writing

Retrieval Skills

For search, RAG, evidence extraction, and knowledge-base answering.

Examples:

• rag-answering
• evidence-based-research
• academic-search
• web-research
• source-grounded-summary

Evaluation Skills

For evaluation, scoring, checking, and quality control.

Examples:

• agent-harness-evaluation-skill
• benchmark-design
• answer-evaluation
• rubric-scoring
• quality-check
• regression-test-analysis

Recommended Skill Template

Use the following template when creating a new skill:

---
name: example-skill
metadata:
  version: "1.0"
  category: "general"
description: "Use this skill when..."
---

# Example Skill

## Purpose

Describe the purpose of this skill.

## When to Use

Use this skill when:

- Condition 1
- Condition 2
- Condition 3

## Inputs

Expected inputs:

- Input 1
- Input 2

## Execution Procedure

### Step 1: Understand the Task

Clarify the goal, constraints, and expected output.

### Step 2: Inspect Available Context

Check existing files, tools, dependencies, documents, or prior state.

### Step 3: Execute the Task

Follow the defined workflow.

### Step 4: Validate the Result

Check correctness, completeness, safety, and formatting.

## Output Requirements

The output should include:

- Required item 1
- Required item 2
- Required item 3

## Constraints

The agent must:

- Constraint 1
- Constraint 2

The agent must not:

- Forbidden behavior 1
- Forbidden behavior 2

## Failure Strategy

If the task cannot be completed:

1. Explain what failed.
2. Explain what was attempted.
3. Provide the best partial result.
4. Suggest the next safe action.

## Quality Checklist

Before finalizing, check:

- [ ] The task goal is satisfied
- [ ] The output follows the requested format
- [ ] No unnecessary implementation is added
- [ ] Existing tools or context were used properly
- [ ] Risks and limitations are stated when relevant

## Examples

Example usage:

```text
Use this skill to...

## Design Principles

Skills in this repository should follow these principles.

### 1. Actionable

A skill should not only describe abstract principles. It should tell the agent:

- When to use it
- What to do first
- What to do next
- When to stop
- What to do when something fails
- What the output should look like

### 2. Reusable

A skill should apply to a class of tasks, not just one conversation.

Not recommended:

```text
Help me fix this login API this time.

Recommended:

When fixing backend API issues, first check routing, HTTP method, parameter binding, security interception, exception logs, and existing project conventions.

3. Composable

A skill should be able to work with other skills.

Example:

Framework-first Coding Skill
+ Minimal Patch Skill
+ Security Review Skill
+ Test Verification Skill

4. Verifiable

A skill should include a checklist so the output can be reviewed.

Example:

- Did the agent reuse existing framework capabilities?
- Did the agent avoid duplicated implementation?
- Did the agent follow project conventions?
- Did the agent explain failure causes when relevant?

5. Token-efficient

A skill should reduce unnecessary output.

Recommended practices:

• Define output formats clearly
• Avoid long background explanations
• Avoid explaining basic concepts repeatedly
• Avoid outputting entire files unnecessarily
• Prefer diffs, focused steps, and minimal changes

Usage

This repository can be used in several ways.

As Claude Code Skills

Place a skill directory in a location recognized by Claude Code.

skills/
└── framework-first-coding/
    ├── SKILL.md
    └── README.md

As Cursor or IDE Rules

Extract the core rules from SKILL.md into a project-level rule file.

Example:

.cursor/rules/framework-first-coding.mdc

As Capability Modules for Custom Agents

In a custom agent system, a skill can be loaded as a capability declaration or strategy module:

Agent receives task
→ Match skill by description
→ Load SKILL.md
→ Follow execution procedure
→ Validate with checklist

As Tool-use Constraints for MCP Agents

For MCP-enabled agents, skills can constrain tool selection, file reading, code search, build checks, test execution, and failure handling.

Maintenance Workflow

When adding a new skill:

1. Define the problem the skill solves.
2. Confirm that it applies to a reusable class of tasks.
3. Create an independent directory.
4. Write SKILL.md.
5. Write README.md.
6. Add example inputs and expected outputs.
7. Add a quality checklist.
8. Register the skill in the root README.md.
9. Iterate based on real failure cases.

Skill Quality Checklist

Before adding or modifying a skill, check:

• [ ] The skill has a clear name
• [ ] The skill has a clear description
• [ ] Applicable scenarios are defined
• [ ] Non-applicable scenarios are defined
• [ ] Execution steps are explicit
• [ ] Output requirements are defined
• [ ] Constraints are defined
• [ ] Failure strategy is included
• [ ] Quality checklist is included
• [ ] The skill reduces repeated work
• [ ] The skill reduces agent behavior uncertainty
• [ ] The skill can be composed with other skills
• [ ] The skill avoids excessive complexity
• [ ] The skill avoids vague principles without execution details

Recommended Combinations

In real agent systems, multiple skills can be combined into a complete workflow.

Coding Agent Combination

framework-first-coding
+ minimal-patch-coding
+ project-convention-following
+ test-verification
+ security-review

Research Agent Combination

web-research
+ evidence-based-summary
+ source-citation
+ contradiction-check
+ report-writing

Document Agent Combination

document-generation
+ outline-planning
+ style-consistency
+ docx-processing
+ final-quality-check

RAG Agent Combination

retrieval-planning
+ query-rewriting
+ evidence-selection
+ answer-grounding
+ hallucination-check

Versioning

Each skill should maintain its own version number.

Example:

metadata:
  version: "1.0"

Suggested versioning:

• 1.0: Initial usable version
• 1.1: Minor rule adjustment
• 1.2: Added examples or checklist items
• 2.0: Major change to execution flow or design purpose

Roadmap

Potential skills to add:

• Minimal Patch Coding Skill
• Tool Failure Strategy Skill
• Dependency Inspection Skill
• Project Convention Skill
• Code Review Skill
• Security Review Skill
• RAG Answering Skill
• Evidence-based Research Skill
• Document Generation Skill
• Agent Loop Control Skill
• Context Compaction Skill
• Human-in-the-loop Skill
• Context Engineering Skill
• Multi-Agent Communication Skill

Summary

This repository manages personally accumulated Agent Skills.

Its core value is:

Turn one-off prompt experience into reusable, composable, maintainable, and verifiable agent capability modules.

By continuously collecting and refining skills, this repository can become a reusable capability library for coding agents, research agents, document agents, MCP agents, and custom Agent OS implementations.