The create-agent skill scaffolds a new LearningAgent by creating the required directory structure, placeholder files, and a Claude Code agent file. It then guides the user through initial configuration.
The create-agent skill MUST be invocable via /learning-agents create <name>. If no name argument is provided, the skill MUST prompt the user for a name.
The skill MUST normalize agent names by converting spaces and uppercase letters to lowercase dashes (e.g., "Rails ActiveJob" becomes "rails-activejob").
Before creation, the skill MUST check .claude/agents/ for an existing file matching <agent-name>.md. If a conflict is found, the skill MUST inform the user and ask how to proceed before taking any action.
If the agent directory .deepwork/learning-agents/<agent-name>/ already exists, the skill MUST inform the user and ask whether to proceed with updating the configuration or stop.
The skill MUST execute the scaffold script at ${CLAUDE_PLUGIN_ROOT}/scripts/create_agent.sh passing the normalized agent name as the first argument and, when a template path is provided, the template path as the second argument.
The create_agent.sh script MUST create the following directory structure when the agent directory does not exist:
.deepwork/learning-agents/<agent-name>/core-knowledge.md-- with TODO placeholder content (when no template is provided; see LA-REQ-002.19 for template behavior).deepwork/learning-agents/<agent-name>/topics/-- with.gitkeep.deepwork/learning-agents/<agent-name>/learnings/-- with.gitkeep.deepwork/learning-agents/<agent-name>/additional_learning_guidelines/-- containingREADME.md,issue_identification.md,issue_investigation.md, andlearning_from_issues.md
The create_agent.sh script MUST create an additional_learning_guidelines/README.md file that describes the purpose of each guideline file (issue_identification.md, issue_investigation.md, learning_from_issues.md).
The create_agent.sh script MUST create .claude/agents/<agent-name>.md with:
- YAML frontmatter containing
nameanddescriptionfields (initially set to"TODO") - Dynamic context injection for
core-knowledge.mdusing!backtick syntax - A Topics section with dynamic listing of topic files
- A Learnings section referencing the learnings directory
The create_agent.sh script MUST replace the __AGENT__ placeholder in the generated Claude agent file with the actual agent name using sed.
The create_agent.sh script MUST NOT overwrite existing directories or files. If the agent directory already exists, it MUST print a message to stderr and skip directory creation. If the Claude agent file already exists, it MUST print a message to stderr and skip file creation.
The create_agent.sh script MUST require a non-empty agent name argument. If no argument is provided, it MUST print a usage message to stderr and exit with code 1.
After scaffolding, the skill MUST ask the user about the agent's domain of expertise and the kinds of tasks it will handle, then update:
core-knowledge.md-- replacing the TODO content with domain expertise written in second person.claude/agents/<agent-name>.mdfrontmatter -- replacing TODOnamewith a human-readable display name and TODOdescriptionwith a concise invocation description (2-3 sentences max)
The skill MUST offer the user the option to seed initial topics or learnings. If the user accepts, the skill MUST create files following the topic and learning file formats defined in LA-REQ-003.
Upon completion, the skill MUST output a summary listing all files created or modified, usage instructions for invoking the agent via the Agent tool, and a note about the learning cycle.
The skill MUST NOT overwrite any existing file without explicit user confirmation.
The create_agent.sh script MUST accept an optional second positional argument [template-agent-path] specifying the path to an existing LearningAgent directory. When not provided, the script MUST behave identically to the non-template flow (creating TODO placeholder content in core-knowledge.md).
When a template path argument is provided, the script MUST verify that the path is an existing directory. If the directory does not exist, the script MUST print an error message containing the path to stderr and exit with code 1.
When a template path argument is provided and the directory exists, the script MUST verify that the template directory contains a core-knowledge.md file. If the file is missing, the script MUST print an error message containing the path to stderr and exit with code 1.
When a valid template path is provided, the script MUST copy core-knowledge.md from the template directory into the new agent's directory instead of creating the TODO placeholder content.
When a valid template path is provided, the script MUST copy all .md files from the template's topics/ directory into the new agent's topics/ directory. If no .md files exist in the template's topics/ directory, the new agent's topics/ directory MUST remain empty (aside from .gitkeep).
When a valid template path is provided, the script MUST copy all .md files from the template's learnings/ directory into the new agent's learnings/ directory. If no .md files exist in the template's learnings/ directory, the new agent's learnings/ directory MUST remain empty (aside from .gitkeep).
When files are copied from a template, the script MUST output a confirmation message for the core-knowledge copy. The script MUST output the count of copied topic files when at least one topic was copied. The script MUST output the count of copied learning files when at least one learning was copied.
When a template was used, the skill MUST read the copied core-knowledge.md and present it to the user, asking whether to keep it as-is, modify it for the new agent's focus, or replace it entirely. The skill MUST also list any copied topics and learnings and offer the user the opportunity to review, remove, or add to them.
The skill MUST parse $ARGUMENTS to extract the agent name (first token) and an optional template path (second token). Both tokens MUST be passed to the scaffold script as separate positional arguments.