openclaw/Comment Forge

Comment Forge

Generate Reddit-native comments that sound like a real person wrote them. Powered by a real Reddit comment corpus and a 7-dimension QA pipeline that catches AI fingerprints.

What It Does

Feed it a post title, body, and existing comments. Get back a natural reply that:

• Matches the thread tone using corpus-informed few-shot prompting
• Passes AI detection via 7-dimension QA scoring (naturalness, value, subtlety, tone, detection risk, length, AI fingerprint)
• Strips AI tells with deterministic anti-AI cleaning (em-dashes, smart quotes, 50+ AI vocabulary swaps)
• Adds subtle humanness with smart typo injection (40% chance, max 1 per draft, never on product names)

Two Modes

Value-First: Pure tactical advice. No product mention. Great for building karma and credibility.

Product-Drop: Mention a product naturally in the reply. Auto-fit scoring determines if the product fits the thread (1-10 score). If it doesn't fit naturally, falls back to value-first.

Pipeline

1. Corpus Sampling - Stratified, score-weighted real Reddit comment examples
2. Fit Scoring - Classify thread intent, recommend mode (optional, for product-drop)
3. Draft Generation - Corpus-informed few-shot prompting via Gemini or OpenRouter
4. QA Pipeline - Score, revise, re-score loop (3 attempts for product-drop, 7 for value-first)
5. Anti-AI Cleaning - Deterministic post-processing strips AI vocabulary, em-dashes, smart quotes
6. Human Touch - Smart typo injection for believable imperfections

Quick Start

bash setup.sh
source .venv/bin/activate

# Value-first (no product)
python3 comment_forge.py --post "Best CRM for small teams?"

# Product-drop
python3 comment_forge.py --post "What tools do you use for email?" \
  --product "Acme Mail" --product-desc "Email automation for small teams"

# With existing comments for tone matching
python3 comment_forge.py --post "How do you handle cold outreach?" \
  --comments "I use Apollo" "LinkedIn works best imo"

# From JSON file
python3 comment_forge.py --file post.json --json

# Skip QA (faster)
python3 comment_forge.py --post "..." --skip-qa

JSON File Format

{
  "title": "Best CRM for small teams?",
  "body": "Looking for something simple...",
  "comments": [
    "I use HubSpot free tier",
    "Notion works if you're small"
  ],
  "product": "Acme CRM",
  "product_url": "https://acme.com",
  "product_description": "Simple CRM for small teams",
  "category": "saas",
  "mode": "product_drop"
}

API Keys

| Key | Required | Purpose | |-----|----------|---------| | GEMINI_API_KEY | Yes (or OpenRouter) | Primary LLM for generation + QA | | OPENROUTER_API_KEY | Fallback | Alternative LLM provider | | CEREBRAS_API_KEY | Optional | Fast fit scoring (free tier) |

QA Dimensions

| Dimension | Weight | What It Checks | |-----------|--------|----------------| | naturalness | 15% | Does it sound like a real person? | | value_contribution | 15% | Does it help the thread? | | subtlety | 20% | Is the product mention (if any) natural? | | tone_match | 10% | Does it match thread + corpus tone? | | detection_risk | 10% | Would redditors flag it as spam? | | length_appropriate | 10% | Right length for this thread type? | | ai_fingerprint | 20% | Em-dashes, AI vocab, perfect grammar? |

Pass threshold: 7.0/10 composite score.