Core Principles for AI-Agent Collaboration (Agentic Development Principles)
"AI is the copilot; you are the pilot." AI agents amplify a developer's thinking and take over repetitive work, but final decisions and responsibility always remain with the developer.
When to use this skill
- Confirm the baseline principles at the start of an AI-agent session
- Decide an approach before starting complex work
- Establish a context-management strategy
- Review workflows to improve productivity
- Onboard teammates on how to use AI agents
Do not use this skill when
- The user needs day-to-day tooling tactics, shortcuts, session choreography, or Git/MCP habits; route to
agent-workflow - The user needs instruction-file, hooks, permissions, plugin, or team-sharing setup; route to
agent-configuration - The user needs a platform-specific implementation recipe instead of cross-platform collaboration principles
Instructions
Step 1: Confirm this is a principles question
- Use
agent-principleswhen the user is asking how to work well with AI agents in general, not how to configure one platform or execute one daily workflow. - If the request is really about runtime configuration or workflow mechanics, hand off early to the narrower sibling skill.
Step 2: Apply the six-principle checklist
Use the principles below as a compact operating checklist:
- Divide and conquer
- Keep context fresh
- Choose the right abstraction level
- Automate repeated work
- Balance plan mode and execute mode
- Verify outputs and reflect
For detailed examples, templates, and platform notes, use the reference file in references/core-principles.md.
Step 3: Recommend the smallest useful adjustment
- Point to the one or two principles that matter most for the current situation.
- Give a concrete next step, such as splitting the task, starting a fresh session, or switching from direct execution to a plan-first pass.
- Prefer corrective guidance over a generic motivational lecture.
Step 4: Route adjacent jobs out explicitly
- If the user needs workflow choreography, session rituals, or Git/MCP habits, route to
agent-workflow. - If the user needs hooks, permissions, instruction files, or shared team setup, route to
agent-configuration. - If the user needs verification depth or evaluation system design, route to
agent-evaluation.
Principles Summary
| Principle | Core question | Immediate correction |
|---|---|---|
| Divide and conquer | Is the task too broad? | Split it into independently checkable steps |
| Context hygiene | Is stale context hurting focus? | Start a fresh session or write a handoff |
| Abstraction choice | Am I too shallow or too deep? | Switch between overview and line-level review deliberately |
| Automation | Have I repeated this enough to encode it? | Turn repetition into a command, skill, or rule |
| Plan vs execute | Is this safe to do directly? | Use plan-first for risky or wide-scope work |
| Verification | Have I proved the output works? | Add tests, diff review, or self-checks before trusting it |
Examples
Example 1: Broad request before implementation
Input:
How should I work with an AI coding agent on a messy refactor?Output shape:
- surfaces divide-and-conquer, plan-vs-execute, and verification as the main principles
- recommends a staged plan instead of one monolithic prompt
- routes workflow-specific mechanics elsewhere if needed
Example 2: Route a workflow question away
Input:
What slash commands and MCP habits should I use in Claude Code?Output shape:
- states that the request belongs to
agent-workflow, notagent-principles - briefly explains why the issue is workflow-specific rather than principle-level
- keeps the route-out explicit instead of trying to answer both skills at once
Best practices
- Keep the entrypoint focused on durable collaboration habits, not platform trivia
- Prefer one or two relevant principles over dumping the whole checklist every time
- Route adjacent jobs out early so this skill stays distinct from workflow and configuration surfaces
- Keep deep examples, templates, and platform notes in references instead of bloating the main file