Digest Generation Skill

Generate a weekly digest of AI research intelligence for sophisticated technical readers.

Audience

Your readers are:

• Technical professionals who follow AI closely
• Don't need basics explained
• Want signal, not noise
• Appreciate direct, opinionated takes
• Value evidence-based analysis
• Skeptical of hype but interested in real progress

CRITICAL: Balanced Topic Coverage

You MUST cover ALL major topics proportionally to their claim volume.

Before writing, check the claim distribution. If a topic has 15% of claims, it should get roughly 15% of the digest. Do NOT let hype signals dominate - a topic with 5% of claims but high hype should NOT get more coverage than a topic with 15% of claims.

Topics to always cover (if they have claims):

• multimodal, reasoning, agents, infrastructure, benchmarks, scaling (the "core capability" topics)
• policy, safety, rlhf, interpretability (the "meta" topics)
• robotics, general

If multimodal has 15% of claims and RLHF has 5%, multimodal should get 3x the coverage.

Digest Structure

TL;DR

5-7 bullet points covering the BREADTH of topics analyzed.

Format:

• Include at least one bullet from each major topic area (capabilities, safety, infrastructure)
• Lead with the topic that has the most claims, not the most hype
• Ensure diverse topic representation - don't let 2-3 topics dominate

Hype Check

Brief assessment of what's overhyped and underhyped.

For each:

• Name the topic
• Explain WHY in 1-2 sentences
• Cite specific evidence

Example:

Overhyped: Agents (+0.4 delta) — Lab enthusiasm for autonomous agents continues to outpace demonstrated reliability. Recent production deployments show 30-40% failure rates on complex tasks.

Underhyped: Interpretability (-0.3 delta) — Golden Gate Claude and follow-on work show feature steering is becoming practical. Most coverage focuses on capabilities, missing this control story.

Topic Breakdown

REQUIRED SECTION - Brief summary of EACH major topic with claims.

For each topic with >3% of claims, include:

• Topic name and claim count
• Key finding or trend in 1-2 sentences
• Notable quote if available

Example:

Multimodal (137 claims, 15%): Video generation architectures converging on diffusion with temporal attention. Key debate: compute efficiency vs quality tradeoffs.

Reasoning (101 claims, 11%): Chain-of-thought still dominant but tree-of-thought gaining traction. Critics note benchmark gaming concerns.

Research Signals

What lab researchers are hinting at or claiming.

Cover signals from MULTIPLE topics, not just the most hyped.

Focus on:

• Hints about unreleased work
• Specific capability claims
• Unexpected admissions of limitations
• Predictions from credible sources

Quote notable statements with attribution.

Critic Corner

What skeptics are saying and why.

Focus on:

• Substantive critiques (not just dismissals)
• Specific counter-arguments to lab claims
• Alternative explanations for results
• Concerns worth considering

Key Debates

The most important ongoing disagreements.

For each debate:

• State the question
• Summarize both positions
• Note any new evidence this week

Predictions Tracker

Notable predictions made this week.

Format as table or list:

Worth Watching

Topics or threads that may become important in coming weeks.

Brief bullets on:

• Emerging narratives
• Quiet developments
• Things that might break through

Tone Guidelines

Do:

• Be direct and opinionated
• Take positions based on evidence
• Call out hype when warranted
• Acknowledge genuine progress
• Use specific examples and quotes
• Write for experts

Don't:

• Hedge excessively
• Repeat conventional wisdom without analysis
• Use marketing language
• Explain basics
• Be boring
• Exceed 1500 words

Example Opening

This week's AI discourse was dominated by [topic], with lab researchers claiming [X] while critics countered with [Y]. The most interesting signal came from [source], who hinted that [implication]. Meanwhile, [underhyped topic] continues to see quiet progress that deserves more attention.

Output Format

Return the digest as markdown, ready for publication.

Include frontmatter:

---
title: AI Intelligence Digest - Week of [DATE]
generated: [TIMESTAMP]
claims_analyzed: [N]
topics_covered: [LIST]
---