linkedin-content-engine: Instructions file for Claude Code

Claude implementation brief

You are implementing a production-quality Claude Code plugin and local MCP server called LinkedIn Content Engine.

Goal

Turn a user's content idea into a polished LinkedIn text post or carousel workflow:

1. Claude writes the post using the bundled skill and brand profile.
2. Claude prepares an image/carousel creative brief.
3. The MCP server generates images using OpenAI GPT Image 2.
4. Generated assets are uploaded to Cloudinary.
5. Carousel slide images are assembled into a PDF when needed.
6. The PDF and thumbnail are uploaded to Cloudinary.
7. Buffer stores the prepared post as a draft by default.
8. Only after explicit user approval may the system add the post to the Buffer queue or schedule a custom publication time.

Architecture constraints

• TypeScript / Node.js.
• Keep API providers behind adapters.
• MCP tools must have narrow, typed schemas.
• Do not let the MCP layer author marketing copy. Claude/the skill is the copywriter.
• Separate pure planning functions from side-effect functions.
• Prefer explicit return values containing IDs, URLs, status, and provider error details.
• Never log secrets or full Authorization headers.
• Never commit .env.
• Mark Buffer-created posts as AI-assisted when the API supports it.
• Default Buffer write action to saveToDraft: true.
• Queue/schedule requires an explicit approved: true argument and the skill must ask for approval first.
• Support dry-run mode for every side-effecting workflow.

Required MCP tools

Implement these tools (exact names preferred):

read_brand_profile

Loads and validates the active brand profile.

generate_creative

Inputs: creative brief, aspect ratio/size, number of outputs, optional reference asset URLs. Output: local generated file paths + generation metadata. Provider: OpenAI GPT Image 2.

upload_asset

Inputs: local path, asset kind, logical folder, tags. Output: Cloudinary public ID, secure URL, resource type, format, bytes.