AgentSkillsCN

context-learning-protocol

自动化协议，用于捕捉、整理并检索项目背景、偏好与经验教训。在重大任务结束时，或在识别出关键用户偏好与项目专属经验教训时使用此技能，以确保这些信息能持久保存于 /context/ 目录中。

SKILL.md
--- frontmatter
name: context-learning-protocol
description: Automated protocol for capturing, organizing, and retrieving project context, preferences, and learnings. Use this skill at the end of significant tasks or when key user preferences or project-specific learnings are identified to ensure they are persistently recorded in /context/.

Context Learning Protocol

Overview

This skill enforces a disciplined approach to knowledge management within the project. It ensures that every significant user preference and technical learning is automatically captured in a structured format under the /context/ directory. This allows the agent to "remember" and "learn" from past interactions, creating a self-improving development environment.

Protocol Workflow

1. Identify Learnings

During or after a task, actively identify TECHNICAL aspects only:

  • •Preferences: User-stated likes/dislikes regarding style, syntax, or workflow (e.g., "User prefers 2-space indentation" or "Always use named parameters").
  • •Learnings: Solutions to tricky bugs, quirks of the codebase, performance optimizations, or discoveries of non-obvious CLI tools and platform-specific utilities (e.g., "Using sips for image processing on macOS" or "Hot reload fails when modifying main.server.dart").

DO NOT save minor UI/styling changes (e.g., color adjustments, padding, font sizes, or layout widths), project management updates, or purely content-related changes. However, ALWAYS capture technical solutions, even if they seem like "one-off" fixes, if they involve a specific tool or methodology that could be reused. Focus on technical context that impacts the development workflow or codebase architecture.

2. File Naming Convention

Save new context files in context/ with descriptive, kebab-case names and a .md extension.

  • •Format: context/[category]-[topic].md
  • •Examples:
    • •context/preference-coding-style.md
    • •context/learning-jaspr-routing.md

3. File Structure

Each context file MUST follow this template:

markdown
# [Title of Context]

## Type
[Preference | Learning]

## Date
[YYYY-MM-DD]

## Description
[Concise description of the context, or finding.]

## Context/Reasoning
[What happened? What is the user's rationale?]

## Actionable Item
[How should this affect future work? e.g., "Always use X for Y", "Do not suggest Z".]

4. Updating the Index

After creating or modifying a context file, YOU MUST update the master index at context/GEMINI.md.

  • •This file contains a table overview of all saved context.
  • •Format for context/GEMINI.md:
markdown
# Project Context Index

| Date | Type | Topic | File |
|------|------|-------|------|
| 2026-02-14 | Learning | Blog Architecture | [learning-blog-arch.md](learning-blog-arch.md) |
| ... | ... | ... | ... |

Automated Triggers

  • •Post-Interaction Check: Before finishing a turn or a task, ask yourself: "Did I learn something permanent?" If yes, execute this protocol.
  • •Explicit Instruction: If the user says "Remember this" or "Save this learning", execute this protocol immediately.

Tool Usage

  • •Use write_file to create new context files.
  • •Use read_file to check context/GEMINI.md before updating it.
  • •Use replace or write_file to update context/GEMINI.md.