Myth AI Digest

A High-Signal Intelligence Filter for Hacker News.

"Don't tell me about the hype. Tell me about the stack." — Myth

Mission

Most AI news is noise. We want signal:

• Engineering Breakthroughs — new architectures, SOTA models
• Agentic Reality — tools that do things (MCP, SWE-bench, computer use)
• Key Figures — what Amodei, Hassabis, or Karpathy are actually shipping

This project runs a daily briefing by loading skills/ai-digest/SKILL.md into an agent runtime.

Architecture

A four-stage pipeline. Each stage is a self-contained prompt with embedded input/output schemas.

Scout → Investigator → Pre-Editor → Editor

1. Scout — Fetches HN page 1. Classifies ~30 items by title and points into in_scope and ambiguous_hot. No article fetches.

2. Investigator — For each ambiguous_hot item, reads the HN comment page only. Confirms or rejects. No external fetches.

3. Pre-Editor — Merges confirmed candidates into a pool. Selects at most 10 by editorial significance. Fetches article body or repo README for each. No link following.

4. Editor — Writes the digest in Myth's voice from the Pre-Editor summaries. No fetches.

Topic taxonomy, breakout rules, and significance ranking are defined in skills/ai-digest/prompts/topics.md.

Usage

Load skills/ai-digest/SKILL.md into your agent runtime. The skill is self-contained — no external state, no database. Output is written to digest/YYYY-MM-DD.md.

Roadmap

• Skill scaffolding
• 4-stage hardened pipeline
• Self-contained stage prompts with embedded schemas
• ProductHunt integration
• HuggingFace Trending