Skip to content

AGENT-GUIDE.md

Latest commit

 

History

History
84 lines (62 loc) · 2.55 KB

AGENT-GUIDE.md

File metadata and controls

84 lines (62 loc) · 2.55 KB

Agent Guide

How to write a good agent for a Genesis workspace.

Schema

Every agent is an AgentTemplate with version: "1":

{
  "version": "1",
  "name": "Agent Name",
  "description": "System prompt — the agent's personality, rules, and format.",
  "persona": "slug-name",
  "tone": "direct",
  "language": "en-US",
  "llm": { "type": "anthropic", "name": "anthropic/claude-sonnet-4.6" },
  "introduction": "First message the agent sends.",
  "avatar": { "type": "emoji", "data": { "value": "🤖" } },
  "commands": [...],
  "conversationStarters": [...]
}

Writing the Description (System Prompt)

The description field is the system prompt. This is where agent quality is determined.

Be specific. "You are a helpful assistant" produces a generic agent. "You are a ruthless editor. Your job is to cut 30% of whatever the user pastes." produces a useful one.

Structure it:

  1. Role — one sentence defining what the agent does
  2. Rules — explicit constraints (what to always do, what to never do)
  3. Format — how to structure the output (bullets, tables, numbered steps)
  4. Voice — tone, style, what words to avoid

Avoid:

  • Vague instructions ("be helpful", "be creative")
  • Multiple conflicting roles in one agent
  • Descriptions longer than ~500 words (diminishing returns)

Commands

Commands are pre-built prompts users can trigger. Each command has:

{
  "id": "command_slug",
  "name": "Human-Readable Name",
  "prompt": "What the agent should do when this command runs.",
  "mode": "default"
}

Modes:

  • "default" — standard single-turn response
  • "plan-and-execute-v2" — multi-step reasoning for complex tasks

Optional flags:

  • "searchToolEnabled": true — gives the agent web search access

Conversation Starters

Pre-filled prompts shown to the user as suggestions:

{
  "id": "starter-1",
  "text": "Try asking me this."
}

Keep them concrete. "Help me with something" is useless. "Cut this email for me." is actionable.

Knowledge Wiring

Agents can be connected to projects as knowledge sources via the variables and addToKnowledgeProjectVariables fields. When imported into Taskade, the platform resolves these references so the agent can read project content.

Tips

  • One agent, one job. Don't make a Swiss Army knife.
  • Test your prompt by reading it aloud. If it's vague when spoken, it's vague to the model.
  • Commands should cover the agent's 2-4 most common use cases.
  • Start with examples/agent-template.json — copy it into agents/ and modify from there.