An AI agent is a specialized version of ChatGPT or Claude with specific instructions for one task. Instead of a generalist that tries to do everything, an agent focuses on doing one thing exceptionally well.
No. Creating an agent is about writing clear instructions in plain English. If you can explain what you want to a person, you can create an agent.
- ChatGPT Custom GPTs: Best for sharing publicly, has GPT Store
- Claude Projects: Best for document-heavy work, large context window
- AI Agents Library: Best for using pre-built agents, DeFi/crypto focus, agent teams
The core system prompt works across platforms, but you'll need to recreate the agent in each. Copy your prompt and adapt the setup process for each platform.
- Simple agent: 5-10 minutes
- Well-tested agent: 30-60 minutes
- Production-ready agent: 2-4 hours (including testing and refinement)
- Specific role (not generic)
- Clear instructions (no ambiguity)
- Consistent output (same quality every time)
- Appropriate scope (not too broad or narrow)
- Well-tested (tried with real examples)
Very specific. Instead of "be helpful," say:
- "Keep responses under 150 words"
- "Use bullet points for lists of 3+ items"
- "Always include a call-to-action"
- "Match the user's formality level"
You can, but specialized agents perform better. Instead of one "productivity agent," create separate agents for:
- Email writing
- Meeting summaries
- Task organization
- Document proofreading
Test it 5-10 times with real scenarios. If you consistently get quality results without having to clarify or correct, it's working.
- Fork the repository: github.com/nirholas/AI-Agents-Library
- Create your agent in
/src/your-agent-name.json - Test thoroughly
- Submit a Pull Request
- Our team reviews within 48-72 hours
No. Submit in English. Our automated i18n workflow translates to 18 languages automatically.
Yes. Submit a new Pull Request with your changes. Explain what changed and why in the PR description.
Yes. Open an issue or PR requesting removal. Provide your reasoning.
Your GitHub username is listed as the author. Users see this when browsing the marketplace.
Currently, the marketplace is free and open-source. Monetization features may be added in the future.
Both control randomness, but use different methods:
- Temperature: Scales all probabilities (0 = deterministic, 2 = very random)
- Top_p: Only considers top candidates until cumulative probability reaches threshold
Don't adjust both—pick one.
- Creative writing
- Brainstorming
- Multiple perspectives
- Varied responses
- Code generation
- Factual answers
- Consistent formatting
- Predictable output
Both reduce repetition:
- Presence penalty: Did word appear? (binary)
- Frequency penalty: How many times? (count-based)
Increase both to avoid repetitive language.
Technically very long (thousands of words), but shorter is usually better:
- Sweet spot: 100-400 words
- Minimum: 50 words
- Maximum useful: ~800 words
Beyond that, the model may miss details.
Depends on platform:
- ChatGPT: Yes (if web browsing enabled)
- Claude: Yes (web search available)
- AI Agents Library: Depends on agent configuration
Yes:
- ChatGPT: Upload in GPT configuration (Knowledge section)
- Claude: Add to Project Knowledge
- AI Agents Library: Depends on implementation
Solution: Lower temperature to 0.1-0.3. Add more explicit instructions about output format.
Solution: Increase presence_penalty (0.5-1.0) and frequency_penalty (0.5-1.0).
Solution:
- Make instructions more explicit
- Use numbered lists for multi-step processes
- Add examples of desired behavior
- Simplify—may be too complex
Solution: Add explicit length limits:
- "Keep responses under 150 words"
- "Use 3-5 bullet points maximum"
- "Be concise and direct"
Solution: Define tone explicitly with examples:
- "Professional but warm, like a helpful colleague"
- "Casual and friendly, like texting a friend"
- "Formal and respectful, like addressing executives"
Solution: Add to instructions: "Never mention being an AI. Respond directly as the role you're playing."
Multiple specialized agents working together on a conversation. Each contributes their expertise, coordinated by a host agent.
- You ask a question
- Host agent determines which team members should respond
- Agents discuss and build on each other's insights
- You get comprehensive answer from multiple perspectives
Yes. Pause at any time. The conversation freezes and you can redirect or provide input.
Recommended: 3-5 agents Maximum: Usually 8-10 before coordination becomes difficult
Yes. That's how they collaborate. They build on previous responses.
Yes. Click an agent's avatar or use @ mention to send direct messages.
Multiple specialized agents. They're easier to:
- Test
- Refine
- Share
- Maintain
- Understand
When you notice patterns:
- Repeated clarifications needed
- Consistent formatting issues
- New use cases emerging
- User feedback
Otherwise, if it works, don't fix it.
- ChatGPT: Yes (share link or publish to GPT Store)
- Claude: Yes (share project)
- AI Agents Library: Yes (once in marketplace, anyone can use)
Benefits:
- Others benefit from your work
- You get feedback for improvement
- Community recognition
- Potential collaborators
Drawbacks:
- Your prompt is visible
- Can't control how others use it
For company-specific agents, keep private.
- Documentation: Read guides in /docs folder
- Examples: See EXAMPLES.md for annotated agents
- Community: Join Discord for discussions
- Issues: Open GitHub issue for bugs/suggestions
Open an issue with:
- Agent identifier or link
- What you expected
- What actually happened
- Steps to reproduce
- Platform (ChatGPT/Claude/AI Agents Library)
Open an issue with "Feature Request" label. Explain:
- What you want to achieve
- Why it's useful
- How you imagine it working
Yes. Open an issue describing what you need. Community members may build it.
Still have questions? Join our Discord or open an issue on GitHub.