Classify Media Relevance

Classifying Media Relevance

Release pages contain two kinds of media: editorial content that belongs in the release (screenshots of the feature, demo videos, diagrams explaining a change) and site chrome that doesn't (author avatars, nav logos, tracking pixels, decorative separators). This skill governs which items end up in a release's media[] array.

The goal is precision-over-recall: a dropped editorial image is recoverable (users click through to the source page), but a kept junk image pollutes the UI and wastes storage.

When this runs

• During the parse pipeline, after the AI extracts release content from a fetched page.
• During crawl-mode fetches, when the extractor reads full-page markdown from a linked article and produces a fresh media[].
• Not during feed fetches where the feed already scoped media to per-entry content (trust the feed).

Cheap pre-checks (keep in code, don't spend AI tokens)

These checks are deterministic, free, and catch the overwhelming majority of obvious junk. Always run them before invoking this skill. If a pre-check drops an item, no AI call is needed.

1. Tracking domains — URL host matches a known tracking/analytics domain (px.ads.linkedin.com, t.co, www.facebook.com/tr, analytics.twitter.com, bat.bing.com). Drop with reason tracking domain: <host>.
2. Unsupported content-type — after HEAD/GET, content-type isn't in the uploadable set (image/png|jpeg|gif|webp|svg+xml|avif, video/mp4|webm). Drop with reason unsupported type.
3. Size bounds — body < 5 KB (tracking pixels, spacers) or > 10 MB (won't upload anyway). Drop.
4. Streaming embeds — YouTube, Vimeo, Loom URLs are kept as type: "video" references without downloading. Never route through R2 upload or this skill.
5. ETag / content hash seen before — if the R2 key derived from content hash already exists, reuse it and skip reclassification.

Everything else — the ambiguous middle where URL patterns overlap between chrome and content — goes through the skill.

Heuristic nudges (optional, low-confidence)

The old code treated path substrings like /avatar, /logo, /icon, /badge, /favicon, 1x1 as hard drops. That was wrong often enough to matter: a post titled "New icon set" shipped images under /icons/ that were the actual product. Do not hard-drop on path substrings. Pass them through as weak negative signals and let the classifier weigh them against context.

The one exception: /favicon.ico and exact /favicon* at the site root are always chrome. Keep that single check in code.

Classification rules

For each remaining media item, decide keep or drop based on these signals, in rough order of importance:

Strong keep signals

• Image or video appears in the middle of release body content (not header/footer of the page).
• Alt text describes a feature, UI state, code, or demo ("New dashboard showing filters", "Architecture diagram", "CLI output").
• Filename suggests editorial content (screenshot-*, demo-*, feature-*, *-hero.png, version numbers in name).
• Dimensions consistent with screenshots/diagrams (wider than 400px, aspect ratio not 1:1 perfect square).
• Hosted on the org's CDN under a posts/releases/blog path (e.g., cdn.example.com/posts/2026/new-thing.png).

Strong drop signals

• Alt text is a person's name, a company name alone, or empty and the URL contains avatar|profile|author|contributor.
• Filename is generic site chrome (logo.svg, wordmark.png, header-bg.jpg, footer-icon.svg).
• Perfect 1:1 square under 200×200 with no contextual link to release content (likely avatar/badge).
• URL path includes /wp-content/plugins/ or /_next/static/media/ with no posts path — usually framework chrome.
• Appears in every release on the source (detectable by callers passing a frequency hint) — site-wide chrome bleeding into parses.

Weak / context-dependent

• /icon, /icons/ paths — only chrome if the release isn't about icons; keep if the release announces icon/design updates.
• /badge, /badges/ — drop if it's a shields.io CI badge, keep if the release is about achievements/credentials.
• SVGs at the top of the page — usually logos, but can be diagrams. Use surrounding alt text and position.

Output format

Return a JSON array, one entry per input item, in the same order:

[
  {
    "url": "https://...",
    "decision": "keep",
    "confidence": "high",
    "reason": "screenshot of new dashboard, alt text describes feature"
  },
  {
    "url": "https://...",
    "decision": "drop",
    "confidence": "high",
    "reason": "author avatar, 80x80 square at top of post"
  }
]

confidence is high when signals align, low when it's a judgment call. Callers treat low drops conservatively — they may keep low-confidence drops on high-value sources.

Anti-patterns

• Don't build a new substring blocklist inside the skill — that's what we're replacing.
• Don't drop based on URL alone without considering alt text, position, and release context.
• Don't request the image bytes to classify — work from URL + alt + surrounding content only. The byte-level decisions happen in the cheap pre-checks.
• Don't keep "just in case" — over-keeping pollutes the grid view more than under-keeping hurts individual releases.