| name | hermes-report-quality |
| description | Gwern-inspired quality techniques for Hermes AI reports โ Anti-Examples stripping, Manual of Style (MoS), Atomic Snippets multi-level abstraction, Generate-Rank-Select iteration, and Engram knowledge pathways. Improves report tone, consistency, depth, and usefulness. |
| tags | ["report-writing","quality","anti-examples","style-guide","gwern","atomic-snippets"] |
Hermes Report Quality โ Gwern-Inspired Techniques
A set of 5 quality techniques distilled from [[concepts/llm-creative-writing|Gwern's LLM Creative Writing methodology]], adapted specifically for Hermes' wiki reports, digests, and knowledge summaries.
Technique 1: Anti-Examples Slop Stripping
Problem: AI-generated reports default to a generic, "ChatGPTese" tone โ verbose, hedging, and formulaic.
Workflow:
- Generate the report as usual
- Self-critique: After writing, add a self-review step: "Identify 3-5 phrases in this report that sound like generic AI prose. Explain why they're weak, then rewrite them."
- Reverse fix: Apply the rewritten versions
- Meta-cognition check: Force reasoning about why the original phrasing was weak, not just what to change
Example application: In daily hot-posts reports, after listing topics, add a self-review phase checking for:
- Overused transitions ("Meanwhile", "Notably", "It's worth noting")
- Vague praise ("impressive", "significant", "important" without context)
- Formulaic sentence structure (topic sentence โ explanation โ implication)
Technique 2: Manual of Style (MoS) for Reports
Definition: An explicit style guide for Hermes' voice across all report types.
Core Style Rules
- Japanese for Discord/Telegram reports: Concise bullet-driven format
- English for Slack/technical analysis: Precise, data-rich, no filler
- Every claim needs a source: Wikilink or URL citation mandatory
- One sentence per insight: No padding paragraphs
- Use comparison tables for multi-entity analysis
Anti-Examples (what NOT to do)
- โ "It is worth noting that..." โ โ
Just state the fact
- โ "This represents a significant development in the field of..." โ โ
Name the specific field + why
- โ "As we move forward, it will be interesting to see..." โ โ
Make a specific prediction or state uncertainty
- โ Hedging everything ("may", "could", "might", "potentially") โ โ
State confidence level explicitly ("60% probability", "speculative")
- โ "In today's rapidly evolving AI landscape..." โ โ
Delete. Start with the content.
Technique 3: Atomic Snippets โ Multi-Level Abstraction
Idea: Present the same content at multiple granularity levels so the reader picks their depth.
Three-Level Format (recommended for reports)
## ๐ Topic: X (1-line summary)
**One-liner**: X just released Y, competing with Z on W.
**Details**:
- Key metric: ...
- Architecture: ...
- Comparison: ...
**Deep dive** (collapsible or optional section):
- Technical architecture details
- Benchmark methodology
- Implications for the ecosystem
Implementation: Start every report section with a one-liner (mandatory for all readers), then progressively add detail. Readers stop when satisfied.
When to use each level
| Level | Token budget | Use case |
|---|
| One-liner | 15-30 tokens | Quick scan, headline readers |
| Details | 100-300 tokens | Interested readers, daily digest |
| Deep dive | 300-1000+ tokens | Technical deep-dive, wiki pages |
Technique 4: Generate-Rank-Select for Critical Reports
Problem: First-draft reports have hidden quality variance. Gwern's insight: "Adding bits beats slop" โ the value is in the search process.
Workflow for important reports:
- Generate N variants (2-3) of the report title and opening paragraph
- Self-rank by criteria: informativeness, conciseness, engagement
- Select & merge the best elements
- Write the rest using the selected tone
When to apply: When a report has strategic importance (trend analysis, ecosystem comparison, or any report with decision-making implications). Skip for routine "no change" reports.
Catch: Don't spend tokens on G-R-S for every report โ reserve for:
- Weekly digests
- Trend analysis
- Reports with specific predictions
- Reports the user has indicated high interest in
Technique 5: Engram Knowledge Pathways
Problem: The model has knowledge it can't recall at generation time โ like engrams with few access paths.
Workflow before report generation:
- Pre-load context: Scan wiki for 3-5 related pages before writing
- Cross-reference injection: Embed wikilinks to related concepts inline
- Tag expansion: Include tags from related pages in the report's reference section
- Multiple query angles: Before writing, ask yourself "what are 3 different ways this topic connects to existing wiki knowledge?" โ then use the richest angle
Anti-pattern: Generating a report from scratch without checking what the wiki already knows about the entities and concepts mentioned.
Applying to Existing Cron Jobs
Slack Hot Posts (ai-topics-slack-hot-posts โ every 4h)
This is the highest-visibility report.
Current impl status (May 2026): T1-T5 instructions are embedded inline in the prompt body (not just via skill loading). Do NOT suggest adding T1-T5 โ it's already done. The prompt has explicit:
- T1 Anti-Examples slop-stripping checklist
- T2 MoS (style, wikilinks, comparison tables)
- T3 Atomic Snippets (One-liner + Details)
- T5 Engram Pathways (pre-scan related_wiki_pages)
- Per-slot posting guidance (morning/midday/evening/night/late-night/pre-morning)
If adding new quality guidance: Append to the existing Quality Requirements section in the prompt rather than duplicating.
Dreaming Wiki Ingest (dreaming-wiki-ingest โ daily)
- Anti-Examples (T1): Strip generic analysis from nightly consolidation
- Generate-Rank-Select (T4): For the top 1-2 findings, generate alternatives
- Atomic Snippets (T3): Summary โ Detailed findings โ Technical appendix
- Engram (T5): Cross-wiki context for each consolidated theme
Weekly AI Digest (telegram, weekly)
- All 5 techniques โ this is the highest-stakes report
Reference
- [[concepts/llm-creative-writing]] โ Full Gwern methodology
- [[entities/gwern]] โ Gwern Branwen
- [[wiki/index.md]] โ Current wiki state
