How I Keep Up with AI Without Drowning in It (Free Skill)

Study Guide

About This Video

A 14-minute video from Cole Medin on how he keeps up with AI news without reading all of it. He shows the AI content engine he has run inside his second brain for over six months, then walks through the free skill that builds your own version.

Summary

The engine tracks hundreds of sources (Reddit, YouTube, RSS feeds, Hacker News, GitHub repositories, changelogs such as Claude Code's) and produces one short daily digest of only what matters to him. It stays cheap and fast because code does the easy work first and models only see what is left. The skill, /build-signal-engine, interviews you about your sources and priorities, then builds the pipeline in tested stages instead of one shot.

The Pipeline, Step by Step

  1. Pull sources (00:42, 03:02). Direct APIs, CLIs, or MCP servers where they work, plus Apify for platforms like YouTube and Reddit where APIs are restricted. Apify returns markdown trimmed for agents.
  2. Deduplicate (03:49). Every item is stored in a central database so nothing is processed twice, which would waste tokens.
  3. Deterministic filters (04:25). Plain code drops stale posts and items with very low engagement, such as a video with a couple hundred views.
  4. Jev first pass (04:59). A cheap model scores each survivor. Obvious fluff exits, obvious keepers pass, and only unsure items continue.
  5. LLM pass (06:11). Claude Haiku handles the uncertain items and prepares the digest. Medin notes Haiku 5.5 gives strong reasoning for the price.
  6. Daily digest (01:04). A concise, readable summary of only what he cares about.

Why Jev Goes First

  • Before Jev, he made an LLM call for every item that passed the code filters.
  • He also tried Jev as the last line of defense with no LLM at all. Accuracy dropped a lot: things he cared about were filtered out and things he did not care about slipped through (05:43).
  • The sweet spot is Jev for the obvious calls and an LLM for the nuanced ones, such as items that need his memory system to judge.
  • Cost example (06:25): a run of 864 posts cost about 3 cents in Jev calls and 42 cents for what reached Claude Haiku, under half a dollar a day.

Three Questions Jev Asks

  • Is this on topic for me?
  • Is this actually new and concrete, or just a toy demo? (07:53)
  • Does it score high enough to keep outright, such as the Sonnet 5.5 announcement or an official 1.0 release of a tool he follows, or should the LLM decide?

His lesson learned: subreddits and blogs are full of toy demos people built to benchmark models, and he wants those dropped immediately.

Strip the Noise Before Any Model Sees It

Do not dump raw MCP or API output into Jev or an LLM. Headers, footers, and extra whitespace can be removed deterministically (06:53). That is one reason he likes Apify, whose output is already agent-ready markdown. All sources live in one configuration file: RSS feeds, YouTube channels, GitHub repositories, arXiv categories, and recurring Reddit, web, and YouTube searches (09:03).

Building Your Own with the Skill

  1. Install (09:47). Follow the README in his skills repository. Claude Code users get a plugin; other agents can be pointed at the repository URL or given just this one skill.
  2. Start fresh (10:11). Run it in a new folder, or give it a path to build elsewhere.
  3. Run /build-signal-engine (10:27). No arguments needed. It first grills you about what you track, what you missed or found late, and how you want to connect to each source. Expect 10 to 15 minutes of answers (11:51).
  4. Six autonomous stages (12:05). It starts with one source end to end, then the rest of the data layer, then code filters, Jev and LLM decision steps, the digest format, and the run schedule. It tests after every stage.
  5. Iterate with the profile (13:19). The skill writes a plain-language profile of your sources, filters, and decision questions. Edit it, then tell the agent to implement the differences. The running engine does not read the profile each time.

Key Takeaways

  1. Use code first (dedupe, staleness, engagement) so models only see what is left.
  2. A cheap classifier for obvious decisions plus an LLM for ambiguous ones beat either alone.
  3. Filter on your own criteria, including "new and concrete, not a toy demo."
  4. Strip boilerplate from source output before any model call.
  5. A skill that interviews you first produces a system fitted to you, unlike a generic open-source repository.

Reflect & Discuss

  • Which of your current news sources produce mostly noise? What simple code rule (age, engagement) would remove it?
  • What would your "on topic" and "new and concrete" questions be?
  • Where in your own workflows could a cheap first-pass model replace a more expensive call?
YouTube