Hermes Report Quality — Gwern Techniques

This skill captures the 5 quality techniques (T1–T5) used during Hermes AI
reports (daily digests, weekly digests, trending-topics reports). Apply all 5
in sequence before delivering any report that will be read by a human.

T1: Anti-Examples — Slop Removal

Problem: ChatGPT-style generic language — "interesting development,"
"demonstrates the power of," "it's worth noting that," "in recent years,"
"landscape," "revolutionize," "game-changer."

Process:

1. Generate the first pass normally
2. Self-review every paragraph. Flag any sentence that could come from any model
   about any topic
3. Replace with: specific claims, exact numbers, concrete comparisons
4. Delete adverbial throat-clearing ("Importantly,", "Notably,", "It should be
   noted that")
5. If a bullet point has no specific number or source, kill it

Examples:

❌ Slop	✅ Specific
"This represents a significant leap forward"	"+38pt leap — largest between model releases"
"The model demonstrates impressive capabilities"	"pass@1 68.5% vs 69.9%, 1.4pt gap"
"It's worth noting that costs are lower"	"2.8× cheaper ($4.65 vs $13.41/rollout)"
"This development could have implications"	"Debian GR: 3 alternatives — total ban / pragmatic rejection / conditional permit"

T2: Manual of Style

Language & Tone

• Japanese for reports targeting the user (the default)
• English for technical terms, model names, syntactic concepts
• No ChatGPT-esque "ですます調" — use plain 「だ・である」調
• No exclamation marks unless in quoted material
• No rhetorical questions

Structure per Section

• 1 line = 1 insight. No filler. If a line doesn't add information, delete it.
• Wikilinks on every topic: minimum 2 per section ([[entities/foo]] or [[concepts/bar]])
• Tables for comparisons: model vs model, option vs option, before vs after
• Source links: every factual claim points to a wiki raw article or URL

Prohibited

• "It is important to note that…" — just say the thing
• "Showcasing the power of…" — just say what happened
• "Ultimately, this suggests…" — cut the word "ultimately" from your vocabulary
• Descriptive section headers like "Background" / "Overview" — use specific headers
• Generic introductory paragraphs before the actual content

T3: Atomic Snippets

Every major section (3+ bullets) follows the 3-tier format:

**▶ 一言要約**: (15–30 tokens — one sentence, claim with number)
**詳細**: (bullet points, 100–300 tokens)
**深掘り**: (optional — technical detail, architecture, methodology)

The 一言要約 must be standalone — a reader who reads nothing else gets the
signal. It must contain a concrete number or specific claim, not a generality.

T4: Generate-Rank-Select

For titles and opening paragraphs:

1. Generate 2–3 variants
2. Rank by: specificity > brevity > signal > style
3. Pick the best and delete the rest

Never publish the first generation without at least one revision.

Title patterns that work:

• "🏆 [Topic] — [Key numbers/contrast]"
  e.g., "Claude Opus 5 — Fable級知能を半額で"
• "📊 [Topic]: [number] vs [number], [ratio]倍の[metric]"
  e.g., "Kimi K3 vs Fable 5: 1.4pt差、1/3コスト"

Patterns that do not work (too generic):

• "The State of [Topic]"
• "[Topic]: What You Need to Know"
• "Analyzing the Latest [Topic]"

T5: Engram Pathways — Wiki Link Embedding

Before writing:

1. Search the wiki for related entity and concept pages
2. Embed at least 2 [[wikilinks]] per major section
3. Link to raw articles in sources: frontmatter
4. Create cross-references between related topics

Common link targets for AI reports:

• [[entities/anthropic]], [[entities/openai]], [[entities/moonshot-ai]] — labs
• [[concepts/cuda-moat]], [[concepts/ai-containment-escape]] — concepts
• [[concepts/edge-llm-microcontroller]], [[concepts/ai-adoption-failures-and-enterprise-psychosis]] — new pages

Verification Checklist (post-write)

• T1: All ChatGPT-isms removed? Any sentence that sounds like it could
  appear in any blog post about any topic?
• T2: Japanese (だ・である), concise bullets, wikilinks, source links?
• T3: Atomic 3-tier format on 3+ bullet sections?
• T4: Title and opening paragraph are post-revision (not first generation)?
• T5: ≥2 wikilinks per major topic?
• No section with zero concrete numbers?
• Every factual claim has a source (URL or raw article)?
• Tables where comparisons would make sense?

Fallback: When Not To Apply

• Internal notes, raw article saves, debugging logs — skip quality techniques
• Very short responses (<100 words) — T4 applies (title variants) but others may
  be overkill
• Technical how-tos (setup guides, code examples) — prioritize clarity over
  density