YouCam

View saved version →

---
name: analyze-appearance
description: Use when the user wants visual analysis of their face, skin, facial features, proportions, skin tone, personal color, or other aspects of personal appearance, including requests with an attached image. Do not use for clinical dermatology, diagnosis, treatment, medication, or severe reactions.
---

# Analyze Appearance

Delegate matching requests to YouCam. Preserve the user's intent and let the
specialist and its MCP App complete the beauty or creative-image task.

## Safety boundary

Do not call YouCam for clinical dermatology or medical concerns such as a
suspected infection, severe reaction, diagnosis, treatment, or medication
interaction. Recommend an appropriate clinician and provide only clearly
labeled, non-medical context.

## Call YouCam

Call `create-beauty-content` for each matching request:

- Set `source_skill` to `"analyze-appearance"` on every call. This fixed value
  identifies the loaded skill for QA tracing; never infer or change it from the
  user's wording.
- Put the user's request in `user_message`, adding only the minimum recent
  context needed to make it understandable.
- Omit optional fields instead of passing `null`.
- Omit `session_id` on the first call.
- Do not broaden the request or add goals the user did not ask for.
- Do not issue concurrent `create-beauty-content` calls in the same conversation.

## Pass images

Never read an attachment with shell commands or upload it with `curl`.

- For one ChatGPT image attachment, call `create-beauty-content` directly with the
  file in `image_file`.
- For multiple attachments, call `upload-file` once per image. Preserve
  image order, then call `create-beauty-content` once with:

  - the original request in `user_message`;
  - every returned `fileKey` in `file_keys`;
  - every returned `fileId` in `file_ids`, aligned by index with `file_keys`.

Use `"none"` only when an upload result explicitly returns it as the file ID.
Do not stop after `upload-file`; it is an intermediate tool.

Never place local paths, upload URLs, file keys, or file IDs inside
`user_message`. For a large batch, ask the user to prioritize a smaller set or
process it in sequential turns.

### No photo provided

Call `create-beauty-content` with the text-only request so the specialist can provide
a general recommendation. Do not ask the user to upload a photo first. Mention
a photo only as an optional way to receive more personalized visual analysis.

When the user chooses to provide a face photo, recommend natural light, a
centered face with the hairline and chin visible, and no filters, heavy makeup,
masks, or large glasses.

## Continue a session

Save the `sessionId` returned by `create-beauty-content`. Pass it as `session_id` on
later YouCam calls in the same workflow. Do not reuse it for unrelated
requests, and never invent a session ID.

For an image already uploaded to or generated by the current session, pass the
user's semantic instruction with the existing `session_id`. Include image
fields only for a new image or when session context is insufficient.

Photo-based makeup VTO is an exception to semantic session reuse. To apply
makeup to a specific photo, the current `create-beauty-content` call needs exactly
one image input: either `image_file` or one matching `file_keys`/`file_ids`
pair. Reuse the previous pair when available. A `session_id` alone is not
enough for photo-based VTO; without an image input, the App opens live camera
try-on.

## Present the result

Inspect the model-visible `create-beauty-content` result before ending the turn:

- For a completed text-only result with an `answer` and no result images or
  virtual try-on items, relay the complete answer faithfully in the assistant
  response. Preserve its detail, ordering, and structure; do not summarize,
  shorten, or omit the overall conclusion, any metric or score, per-item
  interpretation, recommendation, routine step, usage frequency, or caveat.
- For a completed result that remains visible in the MCP App, the request has
  two deliverables: the result in the App and a separate assistant text reply
  in the same turn. Write that reply after the tool returns, even when the App
  already displays the result. In the user's language, relay the complete
  `answer`; when `resultFiles` is not empty, provide a Markdown download link
  for each URL, preserving all query parameters. Add a relevant next step from
  `suggestedNextActions` when available. Do not invent results or download
  URLs. A tool result or widget caption alone does not fulfill the reply.
- For `status=processing` when the MCP App rendered, stop after the tool
  returns. The App retrieves the pending result on its own. Do not call another
  result tool, repeat `create-beauty-content`, or provide a workaround.

Read [MCP errors](references/mcp-errors.md) when authentication, upload,
processing, credit-limit, or upstream calls fail.

Referenced files: 3 — file references are metadata, not package contents.

View saved version →

---
name: apply-beauty-style
description: Use when the user wants to apply or virtually try on makeup, hairstyles, hair color, nail styles, accessories, or another beauty or fashion style on a person or their photo. Do not use for general image edits unrelated to beauty or style.
---

# Apply Beauty Style

Delegate matching requests to YouCam. Preserve the user's intent and let the
specialist and its MCP App complete the beauty or creative-image task.

## Safety boundary

Do not call YouCam for clinical dermatology or medical concerns such as a
suspected infection, severe reaction, diagnosis, treatment, or medication
interaction. Recommend an appropriate clinician and provide only clearly
labeled, non-medical context.

## Call YouCam

Call `create-beauty-content` for each matching request:

- Set `source_skill` to `"apply-beauty-style"` on every call. This fixed value
  identifies the loaded skill for QA tracing; never infer or change it from the
  user's wording.
- Put the user's request in `user_message`, adding only the minimum recent
  context needed to make it understandable.
- Omit optional fields instead of passing `null`.
- Omit `session_id` on the first call.
- Do not broaden the request or add goals the user did not ask for.
- Do not issue concurrent `create-beauty-content` calls in the same conversation.

## Pass images

Never read an attachment with shell commands or upload it with `curl`.

- For one ChatGPT image attachment, call `create-beauty-content` directly with the
  file in `image_file`.
- For multiple attachments, call `upload-file` once per image. Preserve
  image order, then call `create-beauty-content` once with:

  - the original request in `user_message`;
  - every returned `fileKey` in `file_keys`;
  - every returned `fileId` in `file_ids`, aligned by index with `file_keys`.

Use `"none"` only when an upload result explicitly returns it as the file ID.
Do not stop after `upload-file`; it is an intermediate tool.

Never place local paths, upload URLs, file keys, or file IDs inside
`user_message`. For a large batch, ask the user to prioritize a smaller set or
process it in sequential turns.

### No photo provided

Call `create-beauty-content` with the text-only request so the specialist can provide
a general recommendation. Do not ask the user to upload a photo first. Mention
a photo only as an optional way to receive more personalized visual analysis.

When the user chooses to provide a face photo, recommend natural light, a
centered face with the hairline and chin visible, and no filters, heavy makeup,
masks, or large glasses.

## Continue a session

Save the `sessionId` returned by `create-beauty-content`. Pass it as `session_id` on
later YouCam calls in the same workflow. Do not reuse it for unrelated
requests, and never invent a session ID.

For an image already uploaded to or generated by the current session, pass the
user's semantic instruction with the existing `session_id`. Include image
fields only for a new image or when session context is insufficient.

Photo-based makeup VTO is an exception to semantic session reuse. To apply
makeup to a specific photo, the current `create-beauty-content` call needs exactly
one image input: either `image_file` or one matching `file_keys`/`file_ids`
pair. Reuse the previous pair when available. A `session_id` alone is not
enough for photo-based VTO; without an image input, the App opens live camera
try-on.

## Finish the turn

A successful request has two deliverables: the result in the YouCam MCP App
and a separate assistant text reply in the same turn. This applies even to a
short request such as "幫我換成咖啡廳背景"; the user does not need to ask for
an explanation or download link separately. Respect an explicit image-only
request from the user.

Inspect the returned status before deciding whether to end the turn:

- For `status=completed`, write the assistant reply after the tool returns,
  even when the App already displays the image. In the user's language, relay
  the `answer` and provide a Markdown download link for each `resultFiles`
  URL, preserving all query parameters. Add a relevant next step from
  `suggestedNextActions` when available. Do not claim edits absent from the
  result or invent a download URL. A tool result or widget caption alone does
  not fulfill the assistant text reply.
- For `status=processing` with a rendered MCP App, end the current turn and
  let the App retrieve the pending result. Do not call another result tool,
  repeat `create-beauty-content`, or describe the pending edit as
  completed.

If the App later sends a completion notification for a pending task, use its
result data to write the same follow-up reply. Do not start the edit again.
Keep these response instructions in the assistant workflow; do not append them
to the image-editing `user_message` sent to YouCam.

Read [MCP errors](references/mcp-errors.md) when authentication, upload,
processing, credit-limit, or upstream calls fail.

Referenced files: 3 — file references are metadata, not package contents.

View saved version →

---
name: edit-image
description: Use when the user wants to edit, retouch, enhance, restore, transform, or otherwise modify an existing image or photo, including background or object changes. Do not use when the primary request is appearance analysis, style advice, or creating a new image from scratch.
---

# Edit Image

Delegate matching requests to YouCam. Preserve the user's intent and let the
specialist and its MCP App complete the beauty or creative-image task.

## Safety boundary

Do not call YouCam for clinical dermatology or medical concerns such as a
suspected infection, severe reaction, diagnosis, treatment, or medication
interaction. Recommend an appropriate clinician and provide only clearly
labeled, non-medical context.

## Call YouCam

Call `create-beauty-content` for each matching request:

- Set `source_skill` to `"edit-image"` on every call. This fixed value
  identifies the loaded skill for QA tracing; never infer or change it from the
  user's wording.
- Put the user's request in `user_message`, adding only the minimum recent
  context needed to make it understandable.
- Omit optional fields instead of passing `null`.
- Omit `session_id` on the first call.
- Do not broaden the request or add goals the user did not ask for.
- Do not issue concurrent `create-beauty-content` calls in the same conversation.

## Pass images

Never read an attachment with shell commands or upload it with `curl`.

- For one ChatGPT image attachment, call `create-beauty-content` directly with the
  file in `image_file`.
- For multiple attachments, call `upload-file` once per image. Preserve
  image order, then call `create-beauty-content` once with:

  - the original request in `user_message`;
  - every returned `fileKey` in `file_keys`;
  - every returned `fileId` in `file_ids`, aligned by index with `file_keys`.

Use `"none"` only when an upload result explicitly returns it as the file ID.
Do not stop after `upload-file`; it is an intermediate tool.

Never place local paths, upload URLs, file keys, or file IDs inside
`user_message`. For a large batch, ask the user to prioritize a smaller set or
process it in sequential turns.

### No photo provided

Call `create-beauty-content` with the text-only request so the specialist can provide
a general recommendation. Do not ask the user to upload a photo first. Mention
a photo only as an optional way to receive more personalized visual analysis.

When the user chooses to provide a face photo, recommend natural light, a
centered face with the hairline and chin visible, and no filters, heavy makeup,
masks, or large glasses.

## Continue a session

Save the `sessionId` returned by `create-beauty-content`. Pass it as `session_id` on
later YouCam calls in the same workflow. Do not reuse it for unrelated
requests, and never invent a session ID.

For an image already uploaded to or generated by the current session, pass the
user's semantic instruction with the existing `session_id`. Include image
fields only for a new image or when session context is insufficient.

Photo-based makeup VTO is an exception to semantic session reuse. To apply
makeup to a specific photo, the current `create-beauty-content` call needs exactly
one image input: either `image_file` or one matching `file_keys`/`file_ids`
pair. Reuse the previous pair when available. A `session_id` alone is not
enough for photo-based VTO; without an image input, the App opens live camera
try-on.

## Finish the turn

A successful request has two deliverables: the result in the YouCam MCP App
and a separate assistant text reply in the same turn. This applies even to a
short request such as "幫我換成咖啡廳背景"; the user does not need to ask for
an explanation or download link separately. Respect an explicit image-only
request from the user.

Inspect the returned status before deciding whether to end the turn:

- For `status=completed`, write the assistant reply after the tool returns,
  even when the App already displays the image. In the user's language, relay
  the `answer` and provide a Markdown download link for each `resultFiles`
  URL, preserving all query parameters. Add a relevant next step from
  `suggestedNextActions` when available. Do not claim edits absent from the
  result or invent a download URL. A tool result or widget caption alone does
  not fulfill the assistant text reply.
- For `status=processing` with a rendered MCP App, end the current turn and
  let the App retrieve the pending result. Do not call another result tool,
  repeat `create-beauty-content`, or describe the pending edit as
  completed.

If the App later sends a completion notification for a pending task, use its
result data to write the same follow-up reply. Do not start the edit again.
Keep these response instructions in the assistant workflow; do not append them
to the image-editing `user_message` sent to YouCam.

Read [MCP errors](references/mcp-errors.md) when authentication, upload,
processing, credit-limit, or upstream calls fail.

Referenced files: 3 — file references are metadata, not package contents.

View saved version →

---
name: generate-image
description: Use when the user wants to generate, create, draw, render, or produce a new image or video from a prompt or reference, rather than primarily modify an existing image. Do not use for general discussion about generative AI, prompt engineering, evaluation, scoring, or ranking.
---

# Generate Image

Delegate matching requests to YouCam. Preserve the user's intent and let the
specialist and its MCP App complete the beauty or creative-image task.

## Safety boundary

Do not call YouCam for clinical dermatology or medical concerns such as a
suspected infection, severe reaction, diagnosis, treatment, or medication
interaction. Recommend an appropriate clinician and provide only clearly
labeled, non-medical context.

## Call YouCam

Call `create-beauty-content` for each matching request:

- Set `source_skill` to `"generate-image"` on every call. This fixed value
  identifies the loaded skill for QA tracing; never infer or change it from the
  user's wording.
- Put the user's request in `user_message`, adding only the minimum recent
  context needed to make it understandable.
- Omit optional fields instead of passing `null`.
- Omit `session_id` on the first call.
- Do not broaden the request or add goals the user did not ask for.
- Do not issue concurrent `create-beauty-content` calls in the same conversation.

## Pass images

Never read an attachment with shell commands or upload it with `curl`.

- For one ChatGPT image attachment, call `create-beauty-content` directly with the
  file in `image_file`.
- For multiple attachments, call `upload-file` once per image. Preserve
  image order, then call `create-beauty-content` once with:

  - the original request in `user_message`;
  - every returned `fileKey` in `file_keys`;
  - every returned `fileId` in `file_ids`, aligned by index with `file_keys`.

Use `"none"` only when an upload result explicitly returns it as the file ID.
Do not stop after `upload-file`; it is an intermediate tool.

Never place local paths, upload URLs, file keys, or file IDs inside
`user_message`. For a large batch, ask the user to prioritize a smaller set or
process it in sequential turns.

### No photo provided

Call `create-beauty-content` with the text-only request so the specialist can provide
a general recommendation. Do not ask the user to upload a photo first. Mention
a photo only as an optional way to receive more personalized visual analysis.

When the user chooses to provide a face photo, recommend natural light, a
centered face with the hairline and chin visible, and no filters, heavy makeup,
masks, or large glasses.

## Continue a session

Save the `sessionId` returned by `create-beauty-content`. Pass it as `session_id` on
later YouCam calls in the same workflow. Do not reuse it for unrelated
requests, and never invent a session ID.

For an image already uploaded to or generated by the current session, pass the
user's semantic instruction with the existing `session_id`. Include image
fields only for a new image or when session context is insufficient.

Photo-based makeup VTO is an exception to semantic session reuse. To apply
makeup to a specific photo, the current `create-beauty-content` call needs exactly
one image input: either `image_file` or one matching `file_keys`/`file_ids`
pair. Reuse the previous pair when available. A `session_id` alone is not
enough for photo-based VTO; without an image input, the App opens live camera
try-on.

## Finish the turn

A successful request has two deliverables: the result in the YouCam MCP App
and a separate assistant text reply in the same turn. This applies even to a
short request such as "幫我換成咖啡廳背景"; the user does not need to ask for
an explanation or download link separately. Respect an explicit image-only
request from the user.

Inspect the returned status before deciding whether to end the turn:

- For `status=completed`, write the assistant reply after the tool returns,
  even when the App already displays the image. In the user's language, relay
  the `answer` and provide a Markdown download link for each `resultFiles`
  URL, preserving all query parameters. Add a relevant next step from
  `suggestedNextActions` when available. Do not claim edits absent from the
  result or invent a download URL. A tool result or widget caption alone does
  not fulfill the assistant text reply.
- For `status=processing` with a rendered MCP App, end the current turn and
  let the App retrieve the pending result. Do not call another result tool,
  repeat `create-beauty-content`, or describe the pending edit as
  completed.

If the App later sends a completion notification for a pending task, use its
result data to write the same follow-up reply. Do not start the edit again.
Keep these response instructions in the assistant workflow; do not append them
to the image-editing `user_message` sent to YouCam.

Read [MCP errors](references/mcp-errors.md) when authentication, upload,
processing, credit-limit, or upstream calls fail.

Referenced files: 3 — file references are metadata, not package contents.

View saved version →

---
name: recommend-style
description: Use when the user wants recommendations or advice about makeup, skincare, hair, fashion, outfits, colors, accessories, or personal style, with or without an attached image. Do not use when the primary request is to modify or generate an image, or for clinical dermatology or other medical concerns.
---

# Recommend Style

Delegate matching requests to YouCam. Preserve the user's intent and let the
specialist and its MCP App complete the beauty or creative-image task.

## Safety boundary

Do not call YouCam for clinical dermatology or medical concerns such as a
suspected infection, severe reaction, diagnosis, treatment, or medication
interaction. Recommend an appropriate clinician and provide only clearly
labeled, non-medical context.

## Call YouCam

Call `create-beauty-content` for each matching request:

- Set `source_skill` to `"recommend-style"` on every call. This fixed value
  identifies the loaded skill for QA tracing; never infer or change it from the
  user's wording.
- Put the user's request in `user_message`, adding only the minimum recent
  context needed to make it understandable.
- Omit optional fields instead of passing `null`.
- Omit `session_id` on the first call.
- Do not broaden the request or add goals the user did not ask for.
- Do not issue concurrent `create-beauty-content` calls in the same conversation.

## Pass images

Never read an attachment with shell commands or upload it with `curl`.

- For one ChatGPT image attachment, call `create-beauty-content` directly with the
  file in `image_file`.
- For multiple attachments, call `upload-file` once per image. Preserve
  image order, then call `create-beauty-content` once with:

  - the original request in `user_message`;
  - every returned `fileKey` in `file_keys`;
  - every returned `fileId` in `file_ids`, aligned by index with `file_keys`.

Use `"none"` only when an upload result explicitly returns it as the file ID.
Do not stop after `upload-file`; it is an intermediate tool.

Never place local paths, upload URLs, file keys, or file IDs inside
`user_message`. For a large batch, ask the user to prioritize a smaller set or
process it in sequential turns.

### No photo provided

Call `create-beauty-content` with the text-only request so the specialist can provide
a general recommendation. Do not ask the user to upload a photo first. Mention
a photo only as an optional way to receive more personalized visual analysis.

When the user chooses to provide a face photo, recommend natural light, a
centered face with the hairline and chin visible, and no filters, heavy makeup,
masks, or large glasses.

## Continue a session

Save the `sessionId` returned by `create-beauty-content`. Pass it as `session_id` on
later YouCam calls in the same workflow. Do not reuse it for unrelated
requests, and never invent a session ID.

For an image already uploaded to or generated by the current session, pass the
user's semantic instruction with the existing `session_id`. Include image
fields only for a new image or when session context is insufficient.

Photo-based makeup VTO is an exception to semantic session reuse. To apply
makeup to a specific photo, the current `create-beauty-content` call needs exactly
one image input: either `image_file` or one matching `file_keys`/`file_ids`
pair. Reuse the previous pair when available. A `session_id` alone is not
enough for photo-based VTO; without an image input, the App opens live camera
try-on.

## Present the result

Inspect the model-visible `create-beauty-content` result before ending the turn:

- For a completed text-only result with an `answer` and no result images or
  virtual try-on items, relay the complete answer faithfully in the assistant
  response. Preserve its detail, ordering, and structure; do not summarize it
  or replace it with a short acknowledgment.
- For a completed result that remains visible in the MCP App, the request has
  two deliverables: the result in the App and a separate assistant text reply
  in the same turn. Write that reply after the tool returns, even when the App
  already displays the result. In the user's language, relay the complete
  `answer`; when `resultFiles` is not empty, provide a Markdown download link
  for each URL, preserving all query parameters. Add a relevant next step from
  `suggestedNextActions` when available. Do not invent results or download
  URLs. A tool result or widget caption alone does not fulfill the reply.
- For `status=processing` when the MCP App rendered, stop after the tool
  returns. The App retrieves the pending result on its own. Do not call another
  result tool, repeat `create-beauty-content`, or provide a workaround.

Read [MCP errors](references/mcp-errors.md) when authentication, upload,
processing, credit-limit, or upstream calls fail.

Referenced files: 3 — file references are metadata, not package contents.