inkeep-best-practices

📁 inkeep/skills 📅 Jan 24, 2026
1
总安装量
1
周安装量
#47096
全站排名
安装命令
npx skills add https://github.com/inkeep/skills --skill inkeep-best-practices

Agent 安装分布

cursor 1
claude-code 1

Skill 文档

Inkeep Agents SDK Best Practices

Comprehensive guide for building effective multi-agent systems with the Inkeep Agents SDK. Contains best practices across 7 categories, prioritized by impact to guide agent design, tool configuration, and context management.

When to Apply

Reference these guidelines when:

  • Creating new Agents or Sub Agents
  • Configuring MCP servers or function tools
  • Deciding between transfer vs delegate patterns
  • Implementing structured outputs (data/artifact/status components)
  • Managing context via headers and context fetchers
  • Organizing projects and credentials
  • Using the CLI for push/pull workflows

Rule Categories by Priority

Priority Category Impact Prefix
1 Sub Agent Design CRITICAL subagent-
2 Multi-Agent Relationships HIGH relationship-
3 Tools & MCP Servers HIGH tools-
4 Context Management MEDIUM-HIGH context-
5 Structured Outputs MEDIUM outputs-
6 Project Organization MEDIUM project-
7 Development Workflow LOW-MEDIUM workflow-

Quick Reference

1. Sub Agent Design (CRITICAL)

  • subagent-focused-scope – Keep sub agents focused on narrow, well-defined tasks
  • subagent-tool-limit – Limit sub agents to 5-7 related tools
  • subagent-clear-instructions – Write clear, specific instructions in prompts
  • subagent-stable-ids – Use explicit stable IDs for agents across deployments

2. Multi-Agent Relationships (HIGH)

  • relationship-transfer-vs-delegate – Use transfer for handoff, delegate for subtasks
  • relationship-delegate-no-user-response – Delegates cannot respond directly to users
  • relationship-transfer-control – Transfers make receiving agent the primary driver
  • relationship-minimal-complexity – Avoid multi-agent when single agent suffices

3. Tools & MCP Servers (HIGH)

  • tools-mcp-server-config – Configure MCP servers with proper credentials
  • tools-selected-tools – Use selectedTools to limit exposed tools
  • tools-function-tools – Use function tools for custom sandboxed logic
  • tools-credential-references – Reference credentials by ID, resolve at push time

4. Context Management (MEDIUM-HIGH)

  • context-headers – Pass request-time context via headers
  • context-fetchers – Auto-fetch external data at conversation start
  • context-variables – Reference headers/context as {{variables}} in prompts
  • context-artifacts – Save tool call results as artifacts for sharing

5. Structured Outputs (MEDIUM)

  • outputs-data-components – Use data components for rich UI rendering
  • outputs-artifact-components – Define artifact structure with props schema
  • outputs-status-updates – Provide real-time progress during long operations
  • outputs-zod-schemas – Use Zod schemas for type-safe component props

6. Project Organization (MEDIUM)

  • project-logical-grouping – Group related agents/tools in projects
  • project-model-inheritance – Configure models at project level for inheritance
  • project-credential-management – Manage credentials per environment
  • project-stop-conditions – Set stopWhen limits at project/agent level

7. Development Workflow (LOW-MEDIUM)

  • workflow-push-pull – Use CLI push/pull to sync code with Visual Builder
  • workflow-environment-settings – Use createEnvironmentSettings for env-specific config
  • workflow-minimal-config – Don’t add unnecessary properties or fake MCPs

How to Use

Read individual rule files for detailed explanations and code examples:

rules/subagent-focused-scope.md
rules/relationship-transfer-vs-delegate.md
rules/tools-mcp-server-config.md

Each rule file contains:

  • Brief explanation of why it matters
  • Incorrect code example with explanation
  • Correct code example with explanation
  • Additional context and references