Reflection2xhs

Reflection2XHS

Overview

Use this skill when the user wants to turn a local reflection markdown file into a one-page study-note-style poster image. Prefer Markdown over PDF because this workflow depends on semantic structure rather than fixed page layout.

Workflow

1. Read the local reflection markdown file.
2. Extract:

• title
• subtitle / topic line
• 3 to 6 key takeaways
• technical highlights
• one bottom-line summary sentence

1. Build a stable poster prompt for image generation.
2. Generate the poster image with the built-in image generation capability by default.
3. Only use the local OpenAI image API path as an explicit fallback if the user asks for script-only generation and provides their own OPENAI_API_KEY.
4. Save:

• extracted summary json
• prompt text
• generated image in the source file's sibling *_xhs folder

Script

Use [scripts/reflectiontoxhs.py](scripts/reflectiontoxhs.py).

Preview prompt only:

python "C:\Users\20174\.codex\skills\reflection2xhs\scripts\reflection_to_xhs.py" --input "D:\Desktop_Files\World_Model\papers\20_locomotion\AMP\AMP_reflection.MD"

Generate prompt locally, then use built-in image generation:

python "C:\Users\20174\.codex\skills\reflection2xhs\scripts\reflection_to_xhs.py" --input "D:\Desktop_Files\World_Model\papers\20_locomotion\AMP\AMP_reflection.MD"

The generated prompt.txt is then used by the assistant's built-in image generation flow.

When the assistant uses built-in image generation, it should still place the final poster image into the same local output folder as the extracted artifacts, typically:

• _xhs/summary.json
• _xhs/prompt.txt
• _xhs/poster.png

If needed, the local script can also finalize this step by copying the most recent Codex-generated PNG into the output folder:

python "C:\Users\20174\.codex\skills\reflection2xhs\scripts\reflection_to_xhs.py" --input "D:\Desktop_Files\World_Model\papers\20_locomotion\AMP\AMP_reflection.MD" --copy-latest-codex-image

Script-only image generation fallback:

$env:OPENAI_API_KEY="your_api_key"
python "C:\Users\20174\.codex\skills\reflection2xhs\scripts\reflection_to_xhs.py" --input "D:\Desktop_Files\World_Model\papers\20_locomotion\AMP\AMP_reflection.MD" --generate

Notes

• Prefer the built-in image generation capability after Markdown extraction.
• Use the user's own OPENAI_API_KEY only for the script-only fallback path.
• Default model is gpt-image-1.
• Prefer Markdown over PDF. Use PDF only as fallback when Markdown is unavailable.
• Final image output should not remain only under the Codex generated-images cache; copy it into the source markdown's local *_xhs output folder.
• This skill generates a poster prompt and image workflow. It does not preserve the original page layout pixel-for-pixel.

HTML Poster Layout Rules

When generating HTML posters (preferred over AI image generation for text-heavy content):

1. Content-driven sizing: Do NOT hardcode A4 height (height: 297mm). Use width: 210mm only; let height be determined by actual content. No height: 100% on page container.
2. No forced flex stretch: Do NOT use flex: 1 or justify-content: center on section cards — this creates uncomfortable empty interiors. Cards should be content-determined.
3. Line-spacing fill: Use generous line-height (1.7–1.85) to distribute content evenly. Line-height has an upper limit — never exceed 1.85 just to fill space.
4. Two-column grid: Use display: grid; grid-template-columns: 1fr 1fr for the body. Avoid align-items: stretch if it forces card height beyond content.
5. No bottom gap: Bottom section should have margin-top: 0 to avoid dead space between content and footer.
6. No brand keywords in image prompts: Never use "Xiaohongshu" or similar brand names in AI image generation prompts to avoid unwanted watermarks/logos.
7. Section extraction: Support both ## N. title (markdown headers) and N. title (plain numbered) section formats.
8. Natural rhythm: Padding, font sizes, and spacing should create a comfortable reading rhythm. Don't force elements to fill space that content doesn't naturally occupy.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: phanghonghao
• Source: phanghonghao/THU-Awesome-Skills
• License: MIT

Install and usage instructions live in the source repository linked above.