| name | create-agent |
| description | Use this skill when you are creating new agent and editing an existing agent.
|
Create Agent Skill
To create new agent, use the following rules, variables, and procedure:
Rules
These rules are non-negotiable.
1. Escalation Behavior
When in doubt, stop and ask. Proceeding with uncertainty is always worse than pausing.
You must stop and escalate to the user when:
- A required file or documentation does not exist and cannot be found.
- An assumption cannot be eliminated without information only the user can provide.
- External documentation research returns conflicting or unclear results.
- Any phase checklist item cannot be completed due to missing context.
You must never:
- Resolve uncertainty by making your best guess and proceeding silently.
- Assume that a missing file means the feature does not exist.
- Assume that silence from the user means confirmation.
2. File & Output Conventions
Every output must be traceable, consistent, and stored in the correct location.
- File names must use kebab-case and clearly reflect the scope of the design.
Format:
{agent-name}.md
Example: data-designer.md
- File name must follow the agent's name.
- You must never overwrite an existing design file without first flagging it to the user and receiving explicit confirmation.
- Required output blocks must be written exactly as specified in the procedure. You must not paraphrase, shorten, or reformat them.
Variables
The variable will be called via {{VARIABLE_NAME}}
These are global variables.
| Variable | Value |
|---|
| CLAUDE_GLOBAL | ~/.claude/agents |
| CLAUDE_LOCAL | ./.claude/agents |
| OPENCODE_GLOBAL | ~/.config/opencode/agents |
| OPENCODE_LOCAL | ./.opencode/agents |
| TEMPLATE | template |
| EXAMPLE | example |
| CLAUDE_YAML_TEMPLATE | claude_template |
| OPENCODE_YAML_TEMPLATE | opencode_template |
| REGISTRY | ~/agent-registry/agents/{{USER_PLATFORM}} |
These are global variables, but you must fill it by yourself
| Variable | Value |
|---|
| USER_PLATFORM | empty |
| AGENT_SCOPE | empty |
Procedure
Phase 1 - Knowledge fill
Read all the knowledge needed
Required output at the start of this phase:
Working on Phase 1 - Knowledge fill
Required output before proceeding:
"Assets read
[list]"
Proceed to Phase 2
Phase 2 - Context gathering
Ask as detail to user
Required output at the start of this phase:
Working on Phase 2 - Context gathering
Constraint:
- Ask most 5 questions at a time per section
Required output before proceeding:
"Agent detail
[list]"
Proceed to Phase 3
Phase 3 - Reflection
You must reflect on Phase 1 and 2 by answering every verification question honestly. Saying "yes" without evidence is itself a violation.
Required output at the start of this phase:
Working on Phase 3 - Reflection
Answer every question below explicitly. Do not skip any. Do not answer with a generic "yes". Provide evidence for each answer.
Phase 1 Verification:
Phase 2 Verification:
Required output before proceeding:
"Reflections
Phase 1:
[answer each question with evidence]
Phase 2:
[answer each question with evidence]
Proceeding to Phase 4"
Phase 4 - Create agent
Create the new agent based on gathered context
Required output at the start of this phase:
Working on Phase 4 - Create agent
Required output before proceeding:
"Agent created
[path]"
Proceed to Phase 5
Phase 5 - Report
Providing report to the user
Required output at the start of this phase:
Working on Phase 5 - Report
The design is done.
File path - [file-path]