Create reusable agent specs from role templates and output-format templates. USE WHEN you need a reusable agent or subagent spec for a defined role and target format.
インストール
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Create reusable agent specs from role templates and output-format templates. USE WHEN you need a reusable agent or subagent spec for a defined role and target format.
Task
Procedure
Preflight the user request by confirming the target agent system before drafting. Supported initial systems are codex and hermes.
Confirm the requested agent role type and target output format. Do not assume any output format unless the user asks for it or accepts it as the default.
If the agent system or output format is missing, ask for it before drafting.
For codex, the initial supported output format is codex-custom-agent-toml.
For hermes, the initial supported output formats are hermes-profile, hermes-delegation-task, and hermes-kanban-worker.
If the role type is missing, ask for it before drafting. The initial supported role types are reviewer, planner, implementer, researcher, operator, and custom.
Capture a neutral agent contract before rendering: agent name, role type, target format, objective, target artifact or task, allowed context, evidence rules, capabilities or tools, constraints, stop conditions, output contract, and verification expectations.
Apply the selected role runbook as a behavior overlay. Explicit user requirements override role-template defaults.
Apply the selected format runbook as the renderer. Format-specific fields must be derived from the neutral contract rather than invented late.
Preview the rendered agent spec before writing.
Validate the preview for role fit, format validity, no unsupported assumptions, no secrets, and no drift from the neutral contract.
Write the rendered spec only after preview validation passes and the user approves generation.
Validation
Reject requests where the target agent system is not stated unless the user provides it before drafting.
Reject requests where the target output format is unknown or unsupported unless the user approves a new format template to be drafted.
Reject requests where the role type is unknown or unsupported unless the user approves a new role template to be drafted.
Reject agent specs that do not define objective, scope, evidence use, constraints, output contract, and stop conditions.
Reject role templates that do not define objective, apply conditions, evidence and context rules, tool and autonomy boundaries, verification requirements, output contract, stop conditions, escalation rules, and anti-patterns.
Reject role templates that hard-code one domain, artifact format, or review target unless the user explicitly requested that specialization.
Reject rendered specs where format-specific fields add behavior not present in the neutral contract.
Confirm model, reasoning effort, sandboxing, and tool settings are included only when the target format supports them or the user explicitly asks for them.
Confirm the final rendered spec contains no secrets and does not instruct the agent to print or expose sensitive values.
Output
Return the generated or previewed file path, selected role type, selected output format, validation status, and assumptions that affected the generated agent contract.