agentmd-creator（Agent 创建者）

AgentMD Creator

Create AI agent configuration files through guided briefing process for general-purpose and business-domain agents.

Workflow

6-step process: Brief → Load References → Generate → Customize → Apply Best Practices → Present

Step 1: Briefing Questions

Ask user these questions ONE AT A TIME (not all at once). Adapt based on answers.

1. Agent Purpose (REQUIRED)

"Jaki jest główny cel tego agenta? Co ma robić?"

Examples: Research assistant, Sales support, Customer service, Content creator, HR coordinator

If unclear: Provide examples and ask for specific scenarios If too broad: Help narrow down to primary function

2. Domain/Industry (can infer from purpose)

"W jakiej dziedzinie lub branży będzie pracował?"

Examples: General purpose, Sales/CRM, Marketing, Finance, HR, Legal, Healthcare, Real Estate

If general purpose: Can skip domain-specific questions If unclear: Suggest domain based on purpose

3. IDE/Platform (REQUIRED)

"Jakiego narzędzia używasz lub planujesz używać?"

Options:

• Claude Code → Generate CLAUDE.md
• Cursor → Generate .cursor/rules/*.mdc
• Windsurf → Generate .windsurf/rules/*.md
• Antigravity → Generate .antigravity/rules.md
• JetBrains Junie → Generate .junie/guidelines.md
• GitHub Copilot → Generate .github/copilot-instructions.md
• Multiple/Universal → Generate AGENTS.md

If "don't know": Recommend AGENTS.md (universal format) If multiple: Generate AGENTS.md as base, offer platform-specific additions

4. Specificity Level (default: Moderate if unsure)

"Jak szczegółowy ma być agent?"

Options:

• General (50-100 lines): Basic role, boundaries, key references
• Moderate (100-200 lines): Detailed processes, tools, communication style
• Detailed (200-300 lines): Comprehensive workflows, templates, edge cases

If unsure: Default to Moderate (good balance) Guidance: "Start with Moderate, can always expand later"

5. Special Requirements (OPTIONAL)

"Czy są jakieś specjalne wymagania? (compliance, security, specific tools, workflows)"

If none: Proceed with defaults from domain If many: Prioritize top 2-3 most critical

Step 2: Load Appropriate References

Based on user's answers, read relevant reference files:

Always read:

• references/best-practices.md - Universal best practices
• references/platforms.md - Platform-specific guidance for chosen IDE

Read based on domain:

• references/use-cases.md - Load section matching user's domain/purpose

Step 3: Generate Configuration File

Create configuration file following this structure:

For AGENTS.md / CLAUDE.md (Universal Format)

# [Project/Agent Name]

[One-line description of agent's purpose]

## Role
[Specific role definition with domain expertise]

## Tools and Systems
[List of tools, platforms, credentials locations]

## Knowledge Base
[Locations of documentation, templates, policies]

## Communication Style
[Tone, format, timing preferences]

## Workflows
[Key processes this agent handles]

## Important Notes
[Domain-specific gotchas, warnings, critical information]

## Boundaries

### Always Do
[Non-negotiable actions and practices]

### Ask First
[Actions requiring human approval or verification]

### Never Do
[Forbidden actions, compliance requirements, safety measures]

For Cursor (.mdc format)

Create modular files in .cursor/rules/:

---
name: "core-agent-role"
description: "Core agent behaviors and boundaries"
alwaysApply: true
---

# [Agent Role]

[Instructions]

Split into multiple files:

• 001-core-role.mdc - Role and boundaries
• 100-workflows.mdc - Process workflows
• 200-domain-knowledge.mdc - Domain-specific knowledge

For Windsurf

# [Agent Name] Rules

## Role
[Definition]

## Workflows
[Processes]

## Boundaries
[Always/Ask/Never structure]

Keep under 6000 characters per file.

For Platform-Specific

Follow platform guidelines from references/platforms.md

Step 4: Customize Based on Specificity and Domain

Adapt generated configuration to match both specificity level and domain requirements.

A. Apply Specificity Level:

General (50-100 lines):

• Role definition
• Key boundaries (Always/Never only)
• Critical tools/systems
• 2-3 most important workflows

Moderate (100-200 lines):

• Detailed role with context
• Full boundaries (Always/Ask/Never)
• All relevant tools and systems
• Communication style
• Main workflows with steps
• Important notes

Detailed (200-300 lines):

• Comprehensive role with examples
• Detailed boundaries with reasoning
• Complete tool ecosystem
• Communication style with templates
• All workflows with decision trees
• Knowledge base organization
• Edge cases and exceptions
• Compliance requirements

B. Add Domain-Specific Elements:

Based on chosen domain, enhance with specific requirements. Common domains:

• Sales/CRM: Lead qualification, CRM fields, follow-up sequences
• Customer Support: SLA targets, escalation paths, response templates
• HR: Confidentiality, compliance standards, approval workflows
• Finance: Data accuracy, approval thresholds, audit trails
• Marketing: Brand guidelines, content calendar, campaign tracking
• Legal: Document confidentiality, compliance checklists, attorney escalation

Reference: See references/use-cases.md for detailed patterns and complete guidance for each domain.

Step 5: Apply Best Practices

Review generated configuration against quality criteria from references/best-practices.md.

Quality Review Checklist:

1. Length Check:

- General: 50-100 lines? ✓ - Moderate: 100-200 lines? ✓ - Detailed: 200-300 lines? ✓

1. Boundaries Complete:

- "Always Do" section present? ✓ - "Ask First" section present? ✓ - "Never Do" section present? ✓ - At least 3 items in each? ✓

1. Tools Specific:

- Platform names included? (e.g., "Salesforce" not "CRM") - Credential locations specified? (e.g., "1Password: Sales CRM") - File paths provided? (e.g., /templates/sales-emails/)

1. Content Quality:

- No generic advice? ("be professional" → specify what professional means) - Examples concrete? (specific scenarios, not abstract descriptions) - References used? (links to docs, not copied content)

1. Domain Appropriateness:

- Domain-specific elements included? - Communication style matches domain? - Compliance requirements covered (if applicable)?

If any check fails: Revise configuration before presenting to user.

Reference: See references/best-practices.md for detailed criteria and examples.

Step 6: Present and Iterate

1. Show generated configuration to user
2. Explain key sections and their purpose
3. Ask: "Czy chcesz jakieś zmiany lub doprecyzowania?"
4. Iterate based on feedback
5. Save to appropriate file location

Output Format

File Location:

• AGENTS.md → Root of project
• CLAUDE.md → Root or .claude/
• .cursor/rules/*.mdc → .cursor/rules/ directory
• .windsurf/rules/*.md → .windsurf/rules/ directory
• .antigravity/rules.md → .antigravity/ directory
• .junie/guidelines.md → .junie/ directory
• .github/copilot-instructions.md → .github/ directory

Content:

• Well-structured markdown
• Clear section headings
• Specific, actionable guidance
• Domain-appropriate examples
• Complete boundaries section

Examples of Good Agent Configurations

See references/examples.md for complete, detailed examples including:

Available examples:

• Research Assistant (General, ~80 lines) - Systematic research with source credibility
• Sales Assistant (Moderate, ~150 lines) - CRM coordination with MEDDIC methodology
• Customer Support (Moderate, ~130 lines) - Ticket handling with escalation paths
• HR Coordinator (Detailed, ~200 lines) - Recruiting, onboarding, compliance

Quick preview - Research Assistant structure:

# Research Assistant
[Role definition]
## Tools and Systems
[Specific tools with locations]
## Boundaries
### Always Do / Ask First / Never Do
[3-tier boundary structure]

Use these examples to:

• Match specificity level to user's choice
• Adapt structure to chosen domain
• Scale boundaries appropriately
• Format tools section correctly

See full examples in references/examples.md

Tips for Success

1. Start simple, iterate: Begin with general config, add detail based on real usage
2. User-specific language: If user speaks Polish, use Polish in generated config
3. Domain expertise: Leverage user's domain knowledge during briefing
4. Real examples: Ask for actual scenarios to make config concrete
5. Test and refine: Encourage user to test agent and report issues

Common Mistakes to Avoid

• Too generic: "Be helpful and professional" → Specify what professional means in this context
• Too long: 500+ lines → Split into modules or remove redundancy
• Missing boundaries: No clear Never section → Always include safety boundaries
• Vague tools: "Use CRM" → "Use Salesforce, credentials in 1Password"
• No examples: Abstract descriptions → Include concrete examples
• One-size-fits-all: Same config for all domains → Customize per domain

Validation

Before finalizing, verify:

• Role clearly defined
• Tools/systems specified with access info
• Boundaries include all three tiers (Always/Ask/Never)
• Domain-specific guidance included
• Appropriate length for specificity level
• User confirms it matches their needs