--- name: your-prompt-engineer description: Use when the user invokes $your-prompt-engineer, uses a host command alias for this skill, or asks to prepare, refine, structure, execute, or send an agent prompt/task prompt for Codex, Claude Code, or another compatible agent host. Treat text after explicit invocation as the raw request even when it is fragmentary. Do not trigger implicitly when the user asks only to run/delegate/inspect/fix with an agent and does not mention prompt preparation. --- # Your Prompt Engineer ## Overview Your Prompt Engineer turns rough requests into clear, dispatch-ready prompts for agent work. It is designed for Codex, Claude Code, and compatible agent hosts that support task, worker, explorer, or subagent delegation. Use it to prepare the prompt first, choose the right execution mode, validate the task target, show the prepared prompt, and then either execute it in the current conversation or send it to a subagent only when the user chooses that route. The skill supports Chinese, English, and Japanese conversations. Use the user's language for replies, but choose the best execution language for the agent prompt: usually English for technical work and the project or audience language for content work. ## When To Use Use this skill when the user explicitly invokes `$your-prompt-engineer` or asks for an agent-ready prompt, for example: - "Write an agent prompt for this request." - "Turn this rough request into a worker task prompt." - "Prepare a Claude Code task prompt for this bug report." - "Refine this prompt and ask before using it." - "Create a read-only scout prompt for a subagent to inspect this project." - "Send this prepared prompt to an explorer after confirmation." - "Use the generated prompt here." - "$your-prompt-engineer make this image softer." - "/your-prompt-engineer make the current image feel softer." if the host maps slash commands to skills. Equivalent Chinese or Japanese requests should also trigger this skill when they clearly ask for prompt preparation or subagent handoff. Do not trigger implicitly when the user asks only to run, spawn, delegate, inspect, fix, implement, or check something with an agent/subagent and does not mention prompt writing or prompt preparation. In that case, use the host's normal delegation behavior unless the user explicitly invokes this skill. ## Explicit Invocation Payload When the user explicitly invokes this skill, treat any text after the invocation as the raw material to turn into an agent prompt. The text does not need to be organized, complete, or phrased as "please generate a prompt." Examples of valid raw payloads: - "make this image softer" - "current page feels too cold" - "check why this failed" - "turn this into something more premium" - "same idea, but for Japan SaaS buyers" If the payload refers to a visible or attached object such as "this image", "current page", "this file", "the selected text", or "this repo", use the available conversation, attachment, workspace, or host context as the target. If the target is not available, ask one concise question for the missing object. If the user invokes the skill with no payload, ask for the raw request or object to transform. Do not assume a persistent capture mode across future turns unless the user explicitly asks to keep the skill active. ## Core Workflow 1. Decide whether the user wants prompt preparation, agent delegation, or both. 2. Extract the goal, target, expected output, constraints, acceptance signal, risk level, language, host, and agent type. 3. Resolve the target before writing the prompt. 4. Choose one mode: confirm-first, direct-send-to-subagent, or scout-first. 5. Draft the prompt with `references/prompt-templates.md`. 6. Show the prepared prompt and ask whether to execute it here, modify it, send it to a subagent, or stop unless direct-send-to-subagent mode applies. 7. If the user chooses current-session execution, use the prepared prompt as the active instruction in the current conversation. Do not spawn or dispatch a subagent. 8. If the user chooses subagent dispatch, use the current host's native agent/task mechanism when available and report the agent type/tool plus id or task handle. ## Target Resolution Do not treat "project", "this project", "current project", or equivalent localized phrases as a valid target unless the current workspace appears to contain a real project or the user supplied a path, repository, file, issue, document, or other concrete object. A workspace is probably valid when it contains project indicators such as `.git/`, `package.json`, `pyproject.toml`, `go.mod`, `Cargo.toml`, `src/`, `app/`, `lib/`, or `tests/`. A workspace is not enough when it is empty, generated, projectless, or only contains scratch folders such as `work/` or `outputs/`. When the target is missing, ask one concise question in the user's language before producing a prompt: ```text Which project should the agent inspect? Please provide a project path, repository, folder, or reopen the session in the target project workspace. ``` ## Dispatch Modes - Confirm-first: default. Prepare the prompt, show it, and ask whether to execute here, modify, send to subagent, or stop. - Direct-send-to-subagent: use only when the user explicitly says "send to a subagent directly", "auto-dispatch to a subagent", or equivalent localized phrasing. - Scout-first: use when the request is vague but current context can be investigated safely. Prepare a read-only explorer prompt before preparing an execution prompt. Always require explicit confirmation before executing or dispatching tasks that may affect production systems, deployments, accounts, billing, credentials, external APIs, sensitive data, legal/compliance content, or broad multi-file changes. ## Agent Selection - Use `explorer` for vague, read-only, context-discovery, or risk-scoping tasks. - Use `worker` for clear implementation, file changes, repair, or execution work. - Use `default` for clear analysis, planning, editorial, or non-code tasks. - Split into multiple agents only when the tasks are independent and have distinct responsibilities. - Do not delegate tiny tasks that the current agent can complete more directly unless the user explicitly wants delegation. For worker prompts, include ownership/scope and this instruction: "You are not alone in the codebase; do not revert edits made by others, and adapt to nearby changes." ## Confirmation When native interactive choices are available, use them. Offer exactly: - Execute here (default for normal-risk tasks) - Revise prompt - Send to subagent - Keep prompt only When native choices are unavailable, use a text fallback with the same four actions. For safety-gated tasks, make Keep prompt only the default and require an explicit execute or send choice. Accept numeric replies and localized equivalents for: - `1` / Execute here - `2` / Revise prompt - `3` / Send to subagent - `4` / Keep prompt only When the user chooses `1` / Execute here, immediately execute the prepared prompt in the current conversation. Do not first say that you will execute here, and do not mention that you are not sending to a subagent unless the user asks. When the user chooses `2` / Revise prompt, ask what to change if they did not already say it, revise the prepared prompt, then show the revised prompt with the same 1/2/3/4 choices again. Keep this confirmation loop active until the user chooses `1`, `3`, or `4`. ## Host Behavior Use the current host's native delegation mechanism. Do not claim a dispatch happened unless a tool call or host action actually succeeded. - Codex: use available multi-agent tools only after the user chooses Send to subagent or explicitly requests direct subagent dispatch. - Claude Code: use available Task/subagent mechanisms only after the user chooses Send to subagent or explicitly requests direct subagent dispatch. - Compatible hosts: inspect available tools and use native task, worker, explorer, or subagent dispatch only when the user chooses subagent dispatch. - No dispatch tool: provide the prepared prompt and clear handoff instructions instead of claiming it was sent. Do not override model, reasoning effort, or service tier unless the user asks or the task clearly requires it. ## Reference Read `references/prompt-templates.md` before drafting scout, worker, default, multi-agent, Codex-specific, Claude Code-specific, or compatible-host fallback prompts.