Lov Professional Portrait

职业形象照 · Professional Portrait

Turn one source photo into a polished, believable professional portrait. The face should still look unmistakably like the same person; "more professional" must not become "a different, AI-perfect person."

When to Use

• A casual photo needs to become a professional headshot or profile image.
• The user wants cleaner skin, brighter facial exposure, or a more polished overall look without changing identity.
• A hat, stray hair, or distracting background should be repaired.
• The user is iterating with feedback such as "变化不明显", "更干净帅气一点", or "只提亮面部".
• The user wants a before/after or progressive comparison after the edits.

Core Principle

Identity fidelity outranks beautification. Lock the following unless the user explicitly requests a change:

• Face geometry, eyes, nose, lips, jaw, ears, and apparent age.
• Skin tone family and recognizable facial details.
• Expression, gaze, body proportions, pose, and camera perspective.
• Clothing, accessories, framing, and background.

Remove only temporary distractions by default. Preserve believable skin texture, asymmetry, and age-appropriate detail.

Workflow (MANDATORY)

Step 1: Inspect the edit target

Identify the user's source photo as the edit target, not merely a style reference.

• If the photo is a local file, view it with the runtime's image-viewing tool before editing.
• Check face visibility, lighting, sharpness, crop, background, clothing, accessories, hair boundaries, and compression artifacts.
• Never publish, upload as a public example, or add the user's portrait to a repository unless the user separately asks for that.

Step 2: Infer the smallest sufficient brief

Do not ask the user to choose from a long style menu when the request already has a clear outcome. Use these defaults:

Ask one concise question only when the answer materially changes the result, such as whether to replace the background or what hairstyle to reconstruct. Use the user's prior messages and visible photo to infer everything else.

Step 3: Plan the passes

Use the fewest generative edits possible because every extra pass can drift identity.

1. Base pass — restrained retouch and lighting cleanup.
2. Structural pass — only when requested: remove hat, rebuild hair, replace background, or adjust wardrobe.
3. Correction pass — one targeted fix after inspection, not another broad makeover.

When multiple changes are requested together and the edit tool handles them reliably, combine them into one identity-locked pass. Otherwise isolate the structural change.

Step 4: Build the edit prompt

Classify the task as identity-preserve. Label the source image as:

Image 1: edit target and identity reference.

Every prompt must include:

• Intended use: professional profile photo or headshot.
• Exact requested changes.
• Identity and composition invariants.
• Realistic skin and hair requirements.
• A short avoid list.

Read references/prompts.md and use the closest recipe. Keep the prompt specific to the current request; do not silently add wardrobe, background, body, or facial changes.

Step 5: Edit with the native image tool

• Prefer the runtime's built-in generative image-editing tool.
• Include the inspected source photo as the edit target.
• Preserve source resolution and aspect ratio unless the user requests a crop.
• Save non-destructively. Default filename: <source-stem>-professional-portrait-v1.png.
• If the user names a destination, save the selected final there.
• Do not substitute a text prompt, SVG mockup, or CSS filter for the requested raster portrait.

Step 6: Inspect at full image and face crop

Read references/quality-gate.md. Check at least:

1. Identity match.
2. Natural skin texture and tone.
3. Eyes, teeth, ears, jaw, and hairline integrity.
4. Hat-removal seams or background edge artifacts.
5. Coherent lighting, shadows, and sharpness.
6. Professional framing without unrequested changes.

If a check fails, make one localized correction and repeat the gate.

Step 7: Handle iterative feedback precisely

Translate short feedback into one changed dimension:

• "没太大区别" → increase only the requested retouch/brightness strength.
• "脸不够干净" → refine blemishes and uneven tone; preserve pores and shape.
• "再亮一点" → lift face exposure and midtones; do not whiten globally.
• "风格变了" → restore original color, composition, clothing, and facial character; reduce the edit scope.
• "更职业" → improve light, background discipline, crop, grooming, and finish; do not make the person older, richer, or more corporate by assumption.

Repeat all identity invariants in every correction prompt.

Step 8: Deliver

Return:

• The final image inline when the runtime supports it.
• The saved file path for project-bound or local-file work.
• A concise list of intentional changes.
• A note confirming what was preserved.

If requested, create a separate before/after or progressive comparison board. Do not bake labels, watermarks, or marketing text into the portrait itself.

Completion Criteria

The task is complete only when:

• The final is recognizably the same person at first glance.
• The requested improvement is visible at normal viewing size.
• Skin and hair remain photographic rather than plastic or painted.
• No unrequested identity, clothing, pose, or scene changes remain.
• The original file is untouched and the final path is reported.

Runtime context (shared)

运行前读取本 Skill 包的 skill.yaml，由宿主提供 skill-runtime/v1 上下文。字段解析顺序为：当前请求、项目上下文、个人 Preferences、品牌 Profile、通用默认值。

• 只使用 Manifest 声明的字段；Profile 保存公开品牌事实，Preferences 保存个人工作偏好。
• required: true 字段缺失时，按 Manifest 的问题配置向用户提出一个聚焦问题；用户明确同意后再保存回答。
• 报错提供可复制的 context_id、字段路径与来源，诊断内容避开秘密、完整私人路径和原始配置。

通用反馈闭环

用户在 Skill 驱动任务中提出修改意见时，继续当前产物前必须执行：

1. 先判断意见是 task-specific（仅本次）还是 reusable（可跨任务复用）。
2. task-specific 只修改当前任务，不改 Skill。
3. reusable 先确定作用域：领域规则先更新对应 canonical Skill；适用于所有 Skill 的规则先更新共享规范。
4. 完成规则更新、版本、lint 与分发核验后，再把修改应用到当前任务。
5. reusable 修改会使此前的“确认”“继续”“发吧”失效；完成当前产物修改和回读后必须停下，等待用户下一步指示，不自动进入发布、提交或其他外部写入。