Building an AI-Powered Content Pipeline with Claude

The Problem

As someone building a bilingual personal brand (English + Traditional Chinese), I was spending hours every week on a repetitive loop: scan tech news, find interesting topics, draft posts for Threads/LinkedIn/Instagram, then translate for the other language. Most of that loop turned out to be pattern recognition and synthesis, which is what LLMs are good at.

Architecture Overview

The pipeline has two main stages:

1. Ideation Pipeline

# Simplified flow
sources = fetch_all_sources()      # 100+ sources in parallel
items = deduplicate(sources)        # URL hash + Notion dedup
scored = claude_score(items)        # 5-dimension scoring
ideas = generate_angles(scored)     # Bilingual angle generation
write_to_notion(ideas)              # Structured Notion database

The scoring system evaluates each item on 5 dimensions:

• Relevance (0.25 weight): how well it fits my content pillars
• Virality (0.20): social discussion potential
• Timeliness (0.15): how recent and time-sensitive
• Uniqueness (0.20): how novel the angle is
• Brand fit (0.20): can I write from a unique perspective?

2. Creation Pipeline

ideas = fetch_picked_ideas()        # Status = "Picked" in Notion
drafts = generate_drafts(ideas)     # Platform-specific (Threads/LinkedIn/IG)
validated = quality_check(drafts)   # Auto-regenerate if score < 6
final = transcreate(validated)      # EN → ZH-TW trans-creation
write_to_pipeline(final)            # Ready for review

The key distinction: this is trans-creation, not translation. The Chinese version adapts cultural references, adjusts tone for Taiwan’s tech community, and restructures sentences for natural readability.

Key Design Decisions

Claude CLI as Primary Backend

I use Claude’s CLI (via Max subscription) as the primary LLM backend, with the Anthropic API as a fallback. This gives me essentially unlimited usage for scoring and generation without per-token costs.

class ClaudeClient:
    def __init__(self):
        self.use_cli = os.getenv("CLAUDE_USE_CLI", "true") == "true"

    async def generate(self, prompt: str) -> str:
        if self.use_cli:
            return await self._cli_generate(prompt)
        return await self._api_generate(prompt)

Tiered Source Architecture

Not all sources are created equal. The system uses a 4-tier hierarchy:

Tier 4 sources go through keyword filtering against an AI_KEYWORDS set to avoid noise.

What I Learned

1. Scoring beats filtering. Early versions used simple keyword matching. Adding multi-dimensional scoring with Claude dramatically improved content quality.

2. Trans-creation > Translation. Direct translation produces awkward content. Having Claude understand both cultures and adapt the message produces natural-sounding posts.

3. Source health matters. Silent source failures are the worst bugs. I added health tracking that warns after 3 consecutive empty fetches.

4. Cost control is essential. Capping scoring at 300 items per run and using CLI-first keeps costs predictable.

What’s Next

The pipeline currently stops at draft generation. I still review and post by hand. Next on the list: a personal website (you’re reading it!) for longer versions of this content, and eventually automating the publish step.