interior-content-skill

name	interior-content-skill
description	Use when generating Xiaohongshu interior-design marketing content from an existing collector output directory, especially after a separate collector skill has written data/collect-style files.

Interior Content Skill

Overview

This is a thick Claude Code skill for the non-collector half of the interior design content pipeline. It reads a collector output directory, analyzes the target account style, generates image prompts, generates design renderings, and writes Xiaohongshu marketing copy for home-decoration lead generation.

It does not collect Xiaohongshu data. A separate collector skill must produce the input directory.

Input Contract

The input is a collector output directory:

<collect-dir>/
└── <post-folder>/
    ├── metadata.json
    ├── body.txt
    ├── full_text_snapshot.txt      # optional fallback
    ├── image_urls.txt              # optional if images/ exists
    └── images/                     # recommended
        ├── 01.webp
        └── ...

metadata.json should contain title, note_id, liked_count, and collect_count when available. Local images are compressed and converted to data URIs before multimodal analysis.

Required Environment

The skill expects an .env file in this skill directory or the current working directory:

OPENAI_API_KEY=...
OPENAI_BASE_URL=https://api.openai-next.com/v1
IMAGE_GEN_PROVIDER=openai_image
IMAGE_GEN_MODEL=gemini-3.1-flash-image-preview

Never print secret values. Only report whether required keys are set.

First-Time Setup

From the skill directory:

python -m pip install -e .
cp .env.example .env

Then fill .env locally.

Run Pipeline

Use the stable script entrypoint:

python -m scripts.run_from_collect \
  --collect-dir examples/collect-sample/厚来设计 \
  --target-account-id 厚来设计 \
  --area-sqm 180 \
  --layout 四室两厅 \
  --orientation 南北通透 \
  --target-customer 三代同堂改善型家庭 \
  --pain-points 高级感不能冰冷，要兼顾长辈孩子和年轻夫妻互动 \
  --notes "180㎡四室两厅，三代同堂改善住宅。客户希望有静奢老钱风的高级感，但不要酒店样板间式的冰冷；需要开放客餐厨、充足收纳、儿童活动空间，并照顾长辈日常动线。"

Output Contract

Each run writes artifacts to:

data/runs/<request_id>/
├── xiaohongshu_post.md
├── request.json
├── collected_content.json
├── style_dna.json
├── prompt_bundle.json
├── generated_images.json
├── copy_content.json
└── final_post.json

xiaohongshu_post.md is the easiest file to inspect or copy into a demo: it contains the title, body, hashtags, and generated image paths.

Generated images are written to data/generated/ and referenced as /static/generated/<file>.png.

What To Report

After running, report:

• run directory and xiaohongshu_post.md path
• the full final post: title, body, hashtags, and generated image paths
• style_dna.json summary: visual style and copy voice
• prompt_bundle.json positive prompt summary

If a model call falls back because of rate limits or invalid JSON, say which step fell back and whether the final run still completed.

Common Mistakes

• Do not run collector logic here. This skill starts from an existing collector output directory.
• Do not paste or expose .env values.
• Do not use tests/ as the primary interface; use python -m scripts.run_from_collect.
• Do not require the original interior-design-agent repository. This thick skill includes the non-collector runtime code it needs.