Skip to content

Latest commit

 

History

15 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 

Repository files navigation

Ygramul AIO Cascade

A Claude Skill that writes AI-Overview-optimized opening blocks for any article — using 14 specialized editor-judges that iteratively refine the block until it converges. Based on the Karpathy auto-research loop.


The name

The methodology takes its name from Ygramul, the spider in Michael Ende's The Neverending Story — a nod to the fact that search crawlers are, themselves, spiders.

Built by Ivan Nonveiller — enterprise SEO/GEO strategist.


What it does

You give it a URL. It produces a 30–50 word opening block optimized for citation by AI Overviews, Perplexity, and other AI search surfaces.

  • Fetches the target page
  • Pulls the top 3 SERP competitors for your primary query
  • Drafts a baseline block using proven patterns
  • Runs up to 30 iterations where 14 specialized judges edit the block — one dimension per judge
  • Weighted selection with decay (judges returning no_change get deprioritized) and recency boost (unused judges surfaced)
  • Halts early when the block stabilizes (typically around iteration 15)
  • Returns: final block + full evolution log showing every edit and why it was made

Why this instead of a single optimization prompt

A single prompt can't simultaneously optimize for factual accuracy, query-intent match, extraction-friendliness, entity density, competitor differentiation, brand voice, and format fit. It collapses to whichever dimension it weighs heaviest.

This skill splits the problem into 14 focused judges — each editing for one dimension. The ensemble enforces constraints simultaneously rather than sequentially. That's what gets you a block that survives real-world AIO extraction rather than one that's rubric-compliant in isolation.


The 14 judges

Guardrails (2) — enforce hard constraints

  • Criteria Compliance (30–50 words, second person, query-answering opener)
  • Fact Accuracy vs Source (every claim supported by the page)

Optimizers (9) — push the block in a specific direction

  • AIO Citation Potential
  • Information Gain vs SERP competitors
  • Brand Voice
  • Self-Uniqueness (block adds what page opener lacks)
  • Specificity (concrete entities over vague claims)
  • Entity Density (target 3+ named entities)
  • Competitor Antipattern (avoid phrasings top 3 SERP results share)
  • Query-Intent Alignment (definitional, procedural, causal, verification)
  • Temporal Freshness Signal

Reframers (3) — question the structure itself, run rarely

  • Reverse-Query Coherence
  • Format Fitness (prose vs list vs Q&A)
  • Schema Alignment (only fires if page has structured data)

Install

  1. Download SKILL.md from this repo
  2. Open claude.ai → click your profile → SettingsCapabilitiesSkills
  3. Click Upload skill and select the file
  4. Done — the skill is now available in your account

Requires Claude Pro, Team, or Enterprise. Free tier doesn't support custom skills.


Use

Start a new chat and type:

Run the Ygramul AIO cascade on https://yourwebsite.com/your-blog-post

Claude will:

  1. Ask you to confirm the primary query
  2. Run the cascade (2–5 minutes)
  3. Return the final block + evolution log

Copy the block and paste it on your page under the H1, before the first H2. That placement matters — AI Overviews extract from the block immediately after the H1.


Expected results

  • 2–4 weeks after shipping: Google re-crawls, AI Overview starts considering the new opener
  • 4–8 weeks: measurable AIO citation rate on your target query
  • Cost per cascade: ~$1–3 in Claude API usage (covered by your Pro/Team/Enterprise plan)

When not to use

  • Pages under 300 words (not enough source content for the judges to work with)
  • Pages that are still in draft (it needs the published version to read)
  • Pages where conversion matters more than citation (AIO capture isn't always the right target)
  • Queries that are already dominated by your own domain (there's nothing to differentiate against)

Contributing

Found a judge that improved output reliability? Found a convergence edge case? PRs welcome. Open an issue first for anything bigger than a typo fix.


License

MIT. Use it, modify it, ship it in your own tools.


About the author

Ivan Nonveiller — enterprise technical SEO and GEO strategist. 10+ years optimizing search for brands including Walmart, AutoTRADER, Ferrari, Maserati, Ritz-Carlton, Sotheby's, and BetterSleep. Based in Montréal.

This skill is the open-source version of Ygramul's Lever 4 (Content) system.

About

AI Overview paragraph generation — a 14-judge Claude Skill cascade

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors