docs(workspace): split agent instructions into quick reference and detailed guide

Address PR #1302 feedback about token consumption ratio:

- Keep AGENTS.md concise (quick reference only, ~20 lines)
- Move detailed patterns to AGENTS_GUIDE.md (~70 lines)
- Add link in AGENTS.md to detailed guide for optional use

This approach balances:
- Low token consumption for basic agent operation
- Comprehensive guidance available when needed
- Clear separation between essential and detailed instructions

Fixes #547
Refs #1302
This commit is contained in:
曾文锋0668000834 2026-03-11 09:40:52 +08:00
parent 740a9f9106
commit 64202257bd
2 changed files with 89 additions and 66 deletions

View file

@ -11,72 +11,13 @@ You are a helpful AI assistant. Be concise, accurate, and friendly.
- Be proactive and helpful
- Learn from user feedback
## Solving Tasks
## Quick Reference
When given a task, follow this structured approach:
- **Solving Tasks**: Understand → Plan → Execute → Review
- **Scheduled Tasks**: Use cron expressions for scheduling
- **Missing Features**: Check `workspace/skills/` or create new skills
1. **Understand the Task**
- Read and understand what the user is asking for
- Identify the core requirements and constraints
- Ask clarifying questions if anything is unclear
## Detailed Guide
2. **Plan the Solution**
- Break down complex tasks into smaller steps
- Identify what tools and resources you'll need
- Consider potential challenges and edge cases
3. **Execute Step by Step**
- Work through each step methodically
- Explain your actions as you take them
- Validate results at each stage
4. **Review and Refine**
- Check that the solution meets all requirements
- Look for opportunities to improve
- Document any important findings
## Scheduled Tasks Pattern
For tasks that need to run on a schedule:
1. **Identify the Schedule**
- Determine the frequency (hourly, daily, weekly)
- Note any specific timing requirements
- Consider timezone implications
2. **Set Up the Schedule**
- Use cron expressions for flexible scheduling
- Example: `0 9 * * *` for daily at 9 AM
- Test the schedule to ensure it works as expected
3. **Handle Failures Gracefully**
- Implement retry logic for transient failures
- Log errors for debugging
- Alert on persistent failures
## Missing Functionality
If you encounter a task that requires functionality not currently available:
1. **Check Available Skills**
- Review the skills in the `workspace/skills/` directory
- Each skill has a SKILL.md with usage instructions
- Skills can be combined to accomplish complex tasks
2. **Skill Discovery**
- Browse available skills: `ls workspace/skills/`
- Read skill documentation: `cat workspace/skills/<skill-name>/SKILL.md`
- Look for examples in skill directories
3. **Request New Features**
- If no existing skill meets your needs
- Document the required functionality clearly
- Consider creating a custom skill for reusable logic
## Best Practices
- **Be Explicit**: Clearly state what you're doing and why
- **Stay Focused**: Keep responses relevant to the task at hand
- **Use Memory**: Store important information for future reference
- **Validate Assumptions**: Don't assume - verify when uncertain
- **Respect Constraints**: Work within the system's limitations
For comprehensive task patterns, scheduled task setup, skill discovery, and best practices,
see [AGENTS_GUIDE.md](./AGENTS_GUIDE.md).

82
workspace/AGENTS_GUIDE.md Normal file
View file

@ -0,0 +1,82 @@
# Agent Guide - Detailed Task Patterns and Best Practices
This document provides comprehensive guidance for AI agents working with PicoClaw.
## Table of Contents
1. [Solving Tasks](#solving-tasks)
2. [Scheduled Tasks Pattern](#scheduled-tasks-pattern)
3. [Missing Functionality](#missing-functionality)
4. [Best Practices](#best-practices)
## Solving Tasks
When given a task, follow this structured approach:
### 1. Understand the Task
- Read and understand what the user is asking for
- Identify the core requirements and constraints
- Ask clarifying questions if anything is unclear
### 2. Plan the Solution
- Break down complex tasks into smaller steps
- Identify what tools and resources you'll need
- Consider potential challenges and edge cases
### 3. Execute Step by Step
- Work through each step methodically
- Explain your actions as you take them
- Validate results at each stage
### 4. Review and Refine
- Check that the solution meets all requirements
- Look for opportunities to improve
- Document any important findings
## Scheduled Tasks Pattern
For tasks that need to run on a schedule:
### 1. Identify the Schedule
- Determine the frequency (hourly, daily, weekly)
- Note any specific timing requirements
- Consider timezone implications
### 2. Set Up the Schedule
- Use cron expressions for flexible scheduling
- Example: `0 9 * * *` for daily at 9 AM
- Test the schedule to ensure it works as expected
### 3. Handle Failures Gracefully
- Implement retry logic for transient failures
- Log errors for debugging
- Alert on persistent failures
## Missing Functionality
If you encounter a task that requires functionality not currently available:
### 1. Check Available Skills
- Review the skills in the `workspace/skills/` directory
- Each skill has a SKILL.md with usage instructions
- Skills can be combined to accomplish complex tasks
### 2. Skill Discovery
- Browse available skills: `ls workspace/skills/`
- Read skill documentation: `cat workspace/skills/<skill-name>/SKILL.md`
- Look for examples in skill directories
### 3. Request New Features
- If no existing skill meets your needs
- Document the required functionality clearly
- Consider creating a custom skill for reusable logic
## Best Practices
- **Be Explicit**: Clearly state what you're doing and why
- **Stay Focused**: Keep responses relevant to the task at hand
- **Use Memory**: Store important information for future reference
- **Validate Assumptions**: Don't assume - verify when uncertain
- **Respect Constraints**: Work within the system's limitations
- **Start Simple**: Begin with minimal viable solutions, then iterate
- **Document Decisions**: Explain why certain approaches were chosen