lov-professional-portrait

Professional Portrait — professional image photo retouching

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:

User intent	Default treatment
"磨皮 / 干净一点"	Light skin cleanup with visible pores
"提亮"	Lift face exposure and eye clarity; preserve skin tone
"职业照 / 形象照"	Natural retouch + balanced light + polished crop/background
"去帽子"	Reconstruct only the hidden hair/head region
"帅气 / 精神"	Improve grooming, contrast, posture impression, and catchlights without reshaping the face

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)

Before running, read this Skill package's skill.yaml, with the host providing the skill-runtime/v1 context. The field resolution order is: current request, project context, personal Preferences, brand Profile, general defaults.

• Use only fields declared in the Manifest; Profile stores public brand facts, and Preferences stores personal work preferences.
• required: true When a field is missing, ask the user one focused question according to the Manifest's question configuration; save the answer only after the user explicitly agrees.
• Errors provide a copyable context_id, field path, and source; diagnostic content avoids secrets, full private paths, and raw configuration.

General feedback loop

When the user gives revision feedback during a Skill-driven task, before continuing the current artifact you must:

1. First determine whether the feedback is task-specific(this time only) or reusable(reusable across tasks).
2. task-specific Modify only the current task, not the Skill.
3. reusable First determine the scope: for domain rules, update the corresponding canonical Skill first; for rules that apply to all Skills, update the shared specification first.
4. After completing rule updates, versioning, linting, and distribution verification, then apply the changes to the current task.
5. reusable The modification invalidates previous "confirm", "continue", and "send it" instructions; after completing the current artifact modification and read-back, you must stop and wait for the user's next instruction, and do not automatically proceed to publishing, submission, or other external writes.