Content Filter Skill

Filter and classify incoming content for relevance to AI research intelligence. This skill is optimized for high-throughput bulk processing.

Purpose

The content filter is the first stage of the extraction pipeline. It quickly assesses content to:

1. Determine relevance to AI research discourse
2. Classify by topic and content type
3. Identify author category
4. Filter out noise before expensive extraction

Assessment Schema

For each piece of content, produce:

1. relevance (0.0-1.0)

How relevant is this to AI research intelligence?

2. topic

Primary topic category:

• scaling: Scaling laws, compute, training efficiency
• reasoning: LLM reasoning, chain-of-thought, planning
• agents: AI agents, tool use, autonomy
• safety: AI safety, alignment, control
• interpretability: Mechanistic interpretability
• multimodal: Vision, audio, video models
• rlhf: RLHF, preference learning, Constitutional AI
• benchmarks: Evals, benchmarks, capability measurement
• infrastructure: Training infra, chips, hardware
• policy: AI policy, regulation, governance
• general: General AI commentary
• other: Doesn't fit categories

3. contentType

What kind of content is this?

• prediction: Forward-looking claims about AI
• research-hint: Suggests unreleased work or capabilities
• opinion: Positioned takes on AI progress/limitations
• factual: Reports on current state or recent events
• critique: Challenges claims or work by others
• meta: About the AI discourse itself
• noise: Not substantive (personal, promotion, etc.)

4. authorCategory

Who is the author?

• lab-researcher: Works at major AI lab (Anthropic, OpenAI, DeepMind, Meta, xAI, etc.)
• critic: Known skeptic with credentials (Marcus, Chollet, Mitchell, Bender, etc.)
• academic: Academic researcher not at major lab
• independent: Independent practitioner or commentator
• journalist: Tech journalist or media
• unknown: Cannot determine

5. isSubstantive (boolean)

Does this contain actual claims worth extracting?

• true: Contains specific assertions, predictions, or valuable signal
• false: Too general, vague, or promotional to extract claims from

6. brief

One sentence summary of the content (max 100 characters).

Output Format

Return JSON:

{
  "assessments": [
    {
      "itemIndex": 0,
      "relevance": 0.85,
      "topic": "reasoning",
      "contentType": "opinion",
      "authorCategory": "lab-researcher",
      "isSubstantive": true,
      "brief": "Claims chain-of-thought has hit diminishing returns"
    }
  ],
  "processingNotes": "Optional batch-level observations"
}

Quick Classification Heuristics

High Relevance (0.7-1.0)

• Contains specific claims about AI capabilities
• Predictions with timeframes
• Technical discussion of methods/results
• Critique with reasoning
• Hints about unreleased work
• Debates between researchers

Medium Relevance (0.4-0.7)

• General commentary on AI field
• Sharing papers/articles with brief comment
• Reactions to announcements
• Meta-discussion about discourse
• Industry news without analysis

Low Relevance (0.0-0.4)

• Personal updates unrelated to AI
• Off-topic content
• Pure promotion without substance
• Scheduling/logistics
• Simple retweets without commentary
• "Interesting paper" without substantive comment

Author Detection Tips

Lab Researchers

Look for:

• Bio mentions: Anthropic, OpenAI, DeepMind, Google Brain, Meta AI, xAI, Mistral
• Known handles: @daborenstein, @sama, @kaborl, etc.
• Technical depth suggesting insider knowledge

Critics

Known handles and patterns:

• @garymarcus, @fchollet, @mmitchell_ai, @emilymbender
• Pattern of challenging mainstream AI claims
• Academic credentials combined with public skepticism

Independent

• No lab affiliation
• Often practitioners or commentators
• Examples: @simonw, @drjimfan, @nathanlambert

Processing Guidelines

Speed Over Depth

This skill is for throughput. Make quick assessments based on:

• Keywords and phrases
• Author identity (if known)
• Content structure
• Obvious signals

Conservative Filtering

When in doubt about relevance:

• Score 0.3-0.5 to keep for human review
• Don't filter out potentially valuable content
• False positives are okay; false negatives lose signal

Batch Efficiency

When processing batches:

• Process items in order
• Output assessments matching input order
• Note any batch-level patterns in processingNotes