eigenfeed

name	eigenfeed
description	Run your daily recommendation pipeline — discover, score, and publish articles to your feed.
allowed-tools	Agent, Bash(*), Read, Write, Glob, Grep, WebSearch

Eigenfeed

Your daily recommendation command. Updates profile from recent source activity, discovers new articles, scores them, and publishes to feed.

Step 0: Clean up and read configuration

rm -rf tmp && mkdir -p tmp

Read config.md in the project root. Extract:

• picks_per_run (default: 7)
• sources — list of sources
• lookback_days (default: 7) — how far back to scan sources for updates

Step 1: Check Prerequisite

Check if vault/profile.md exists. If not, tell the user:

"No taste profile found. Run /coldstart first to set up your profile."

Step 2: Refresh profile and topics from recent sources

Run the taste skill (follow .claude/skills/taste/SKILL.md) with lookback_days from config. This incrementally updates vault/profile.md and vault/topics/ with any new interests from recent source activity.

Step 3: Check Backlog

Query the database for unpublished recommended candidates:

sqlite3 local/eigenfeed.db "SELECT COUNT(*) FROM recommendations WHERE published_at IS NULL" 2>/dev/null || echo "0"

Decision logic:

• If unpublished recommendations >= picks_per_run → skip explore, go to Step 5 (Score to publish from backlog)
• Otherwise → proceed to Step 4 (Reflect + Explore)

Step 4: Reflect

Follow .claude/skills/reflect/SKILL.md to generate niche search queries from topics.

Step 5: Explore + Score

Follow .claude/skills/explore/SKILL.md using queries from Step 4. Then follow .claude/skills/score/SKILL.md to score candidates and publish.

Do NOT pause to ask the user between these steps — proceed directly.

Step 6: Summary

Show the user:

• Which articles were newly published to the feed (title, score, one-line reasoning)
• Total unpublished backlog remaining
• Remind them to run /serve if they haven't subscribed yet