content-forge

name

content-forge

description

Cross-platform content distillation engine. Collects viral posts from X/Twitter, XHS (Xiaohongshu), and WeChat Official Accounts (GZH), distills top content by engagement scoring, synthesizes original posts via Claude, and learns from human edits to improve over time. Use when the user wants to: (1) generate social media content from trending topics, (2) create cross-platform posts (X threads, XHS notes, GZH articles), (3) run a content pipeline that collects and remixes viral posts, (4) learn from editorial feedback to improve writing style, or mentions "content forge", "content distillation", "cross-platform content", "viral post", or "social content generation".

ContentForge

Cross-platform content distillation and self-learning engine.

Pipeline

A: Collect  →  B: Distill  →  C: Synthesize  →  D: Review  →  E: Learn  →  F: Promote
(bird/xhs/gzh)  (dedup+rank)   (Claude gen)    (human edit)  (style update)  (good→pool)

Prerequisites

• Node.js >= 18
• bird CLI — for X/Twitter search (only needed if platform includes x)
• Python 3 — for XHS/GZH scrapers (only needed if those platforms are used)
• Claude CLI — claude in PATH for synthesis step

Quick Start

# Generate content for a keyword (X only)
node scripts/content-forge-pipeline.js --keyword "AI agents" --platforms x

# All 3 platforms
node scripts/content-forge-pipeline.js --keyword "AI agents" --platforms x,xhs,gzh

# Learn from yesterday's human edits
node scripts/content-forge-pipeline.js --learn

# Check recent runs
node scripts/content-forge-pipeline.js --status

Environment Variables

Variable	Default	Description
BIRD_BIN	bird	bird CLI path
PYTHON_BIN	python3	Python interpreter
CLAUDE_BIN	claude	Claude CLI binary
MIN_FOLLOWERS	1000	Min followers for X source filter
CONTENT_FORGE_DATA_DIR	./content-forge-data	Pipeline data output dir
SOCIAL_MEDIA_DIR	./Social Media	Platform article copies dir

Phase Details

A: Collect

• X: bird search <kw> -n 100 --json-full → filter by followers/language/original-only
• XHS: spawn xhs-search.py --keyword <kw> --limit 30
• GZH: spawn gzh-search.py --keyword <kw> --limit 30

B: Distill

1. Exact dedup (SHA256)
2. Near-dedup (3-gram Jaccard > 0.6, keep higher engagement)
3. Engagement scoring: log10(metric+1) with platform-specific weights
4. Top 30 returned

C: Synthesize

Prompt = style rules + good article examples + distilled sources + anti-AI rules → Claude JSON output (x_thread, xhs_note, gzh_article).

D: Review

Generate .md with >>>original>>> / >>>edit>>> / >>>score(1-5)>>> / >>>comment>>> blocks. Human edits inline.

E: Learn

Parse edits → Claude extracts patterns → auto-update references/style-rules-{platform}.md.

F: Promote

Articles with likes > 50 or shares > 20 → forge-good-articles.jsonl (rolling 20/platform) → feeds back into Phase C.

Style Rules (references/)

Auto-updated by learn phase. Read when customizing output style:

• references/style-rules-x.md — X thread conventions
• references/style-rules-xhs.md — XHS note conventions
• references/style-rules-gzh.md — GZH article conventions

Scripts

• scripts/content-forge.js — Core engine (pure functions, no side effects on import)
• scripts/content-forge-pipeline.js — CLI entry point