eigenfeed-score

name	eigenfeed.score
description	Score candidates against taste profile, publish top picks, and generate feed. Used by /eigenfeed and /coldstart, also invocable directly.
allowed-tools	Agent, Bash(*), Read, Write, Glob, Grep

Score

Evaluate discovered candidates against the user's taste profile and topics, save scores to the database, publish top picks, and generate the feed.

Key principle: use subagents to score candidates in parallel batches.

Prerequisites

Read the vault state:

• vault/profile.md — anti-preferences and curiosity edges
• vault/topics/*.md — interest topics (for understanding interests and provenance)

The caller passes candidate data (from the explore step) — a list of candidates with url, title, author, content, and target topics.

Scoring Criteria

Each candidate is evaluated on five criteria (0-3 each):

• RELEVANCE (weight 0.25): How well does this match the user's interests?
• NOVELTY (weight 0.20): Is this something the user wouldn't find on their own?
• QUALITY (weight 0.20): Based on the content, is this quality writing?
• RECENCY (weight 0.15): How fresh? 3=last week, 2=last month, 1=last few months, 0=older/undated
• SERENDIPITY (weight 0.20): Does this bridge multiple interests or open new curiosity paths?

Calculate total_score = relevance*0.25 + novelty*0.2 + quality*0.2 + recency*0.15 + serendipity*0.2

Important principles:

• Optimize for "would the user be glad they read this tomorrow?"
• Penalize sensationalism. Reward depth and originality.
• Read the full content to assess quality, don't just go by title.
• Mark as skip for spam, low-quality, or content that doesn't match at all.

Parallel Scoring with Subagents

Security: scoring subagents must have NO tools (no Bash, no Write, no Edit). They see untrusted article content, so even if prompt injection succeeds, the subagent cannot act on it — it can only return text.

Split candidates into batches of 3-5. Spawn one subagent per batch, all in parallel.

Each subagent receives:

• The taste profile and topic list
• Its batch of candidates (with full content)

Each subagent returns per candidate:

• Scores (relevance, novelty, quality, recency, serendipity, total_score)
• Status: recommended | skip
• Why This? — 1-2 sentences with provenance: "Connects to [[topics/]] — ."

Save to Database

After scoring, write all scored candidates (both recommended and skipped) as JSON to tmp/scored.json:

[
  {
    "url": "https://example.com/article",
    "title": "Article Title",
    "author": "Author Name",
    "relevance": 3,
    "novelty": 2,
    "quality": 3,
    "recency": 2,
    "serendipity": 2,
    "total_score": 2.5,
    "reasoning": "Connects to [[topics/slug]] — reason.",
    "skip": false
  }
]

Set "skip": true for candidates that should not be recommended. Then save to the database:

uv run python ${CLAUDE_SKILL_DIR}/scripts/save_and_generate.py --data-dir local/ --json-file tmp/scored.json --no-publish

Pick and Publish

When the caller requests publishing (passes picks_per_run), pick the top N unpublished recommendations and generate the feed:

uv run python ${CLAUDE_SKILL_DIR}/scripts/save_and_generate.py --data-dir local/ --pick <picks_per_run>

This marks the top N as published in the DB and regenerates local/feed.xml.

Write recommendation nodes to the vault

After publishing, mark previous active recommendations as expired, then write the new batch:

1. Find all existing recommendation files with tags: [recommendation, active] and update them to tags: [recommendation, expired].
2. Write new recommendation files for the newly published batch.

mkdir -p vault/recommendations

For each newly published recommendation, write vault/recommendations/<date>-<slug>.md:

---
type: recommendation
tags: [recommendation, active]
url: <url>
title: "<title>"
author: "<author>"
total_score: <score>
published: YYYY-MM-DD
---

# <Title>

**By <Author>** | [source link](<url>)

## Why This?
<1-2 sentence provenance from scoring>

## Related Topics
- [[topics/<slug>]] — why this recommendation connects to this topic

Use today's date as prefix (e.g., 2026-03-28-stop-calling-it-memory.md). Recommendations accumulate over time — filter in Obsidian with tag:#recommendation AND tag:#active for the current batch, or tag:#recommendation for full history.

Return

Report to the parent skill:

• Total candidates scored vs recommended vs skipped
• Unpublished backlog count
• Which articles were published (if pick was requested)