How AI Content Adaptation Actually Works

2 min read

Most “AI content tools” just paraphrase the same text five times. Real adaptation means respecting each platform’s constraints — and that requires structured prompts, not just “make this shorter.”

The architecture

Content Pipeline uses a two-stage AI pipeline:

Stage 1: Planning (grok-4.3)

The planner receives the full blog post and decides:

• Which platforms to target
• What tone and style to use per platform
• Whether to schedule or post immediately
• Hashtag strategy

Output is a structured JSON plan, not free-form text.

Stage 2: Adaptation (glm-5.1)

Each platform gets its own processor with:

• A cached system prompt containing brand voice guidelines
• The full post body as context
• A platform-specific brief (character limits, style rules, media requirements)

The processor outputs platform-ready content — X threads as an array of tweets, LinkedIn as formatted paragraphs, TikTok as a scene script.

Prompt engineering for constraints

The key insight: constraints go in the system prompt, content goes in the user prompt.

System: You are a content adapter for X (Twitter).
Rules: Max 280 chars per tweet. Lead with a hook.
Use 1-2 hashtags max. Output as JSON array of strings.

User: Adapt this blog post into a 4-tweet thread:
[full post text]

This separation means the system prompt can be cached by the provider (saving tokens and latency), while the user prompt changes per post.

The human review loop

AI output stops at pending_approval state. Every piece of content is reviewed by a human before publishing. This isn’t optional — it’s the core design principle.

The admin dashboard shows:

• Original post for reference
• AI-generated content per platform
• Edit, approve, or reject actions
• Publication confirmation via verifier

What I learned

1. Separate planning from execution — one model plans, another executes
2. Cache aggressively — system prompts + brand guidelines don’t change
3. Never auto-publish — AI makes subtle mistakes that humans catch instantly
4. Platform constraints are hard rules — encode them in prompts, not post-processing

The result: a post goes from idea to 5-platform distribution in under 5 minutes, with human review at every step.