Tag icon generation: prompt feedback system & preferred prefixes

[madjin]

Summary

We've built scripts/posters/generate-tag-icons.py to generate 1:1 icons from tag words using minimal prompts. Through experimentation, we've learned that archetype-based prefixes work better than abstract meta-instructions.

Key Insight

Instead of telling the model that it's wrong, tell it which human mistake it's making:

• ❌ "Semantically incorrect interpretation of..." (too abstract)
• ✅ "Stock photo misunderstanding of..." (encodes the mistake)

Preferred Prefixes (so far)

Based on testing, these work well:

• Stock photo misunderstanding of {tag} - very reliable
• Overconfident intern interpretation of {tag} - good humor
• Idiomatic literal interpretation of {tag} - strong for verbs
• Visually misread meaning of {tag} - especially for "social"
• Literal but wrong interpretation of {tag} - decent fallback

Available Prefixes

icon, stock, linkedin, childrens, headline, idiomatic, 
wikipedia, intern, pun, celebrity, boomer

Old meta-instruction prefixes are commented out in the code but preserved.

Future: Feedback System

Problem: Hard to systematically track which prompts produce good results.

Idea: Fork one of our Discord bots to create a feedback loop:

• Post generated images to a channel
• React with 👍/👎 to rate
• Track which prefix + tag combinations work best
• Build a "best archetype → tag" mapping over time

Could also:

• Generate a prompt matrix (safe / spicy / unhinged tiers)
• A/B test which misreads perform best
• Auto-retry failed tags with different prefixes

Related Files

• scripts/posters/generate-tag-icons.py - main script
• scripts/posters/tag-icons/ - output directory
• scripts/posters/README-posters.md - documentation

Usage

# Single prefix
python scripts/posters/generate-tag-icons.py "security" -p stock

# Random rotation
python scripts/posters/generate-tag-icons.py --from-facts -p random

# List all prefixes
python scripts/posters/generate-tag-icons.py --list-prefixes