AI:CML 1.0 — AI Content Markup Language

AICML will also be known as LLX (Large-Language eXchange).

Votes 0

AI:CML (LLX) 1.0 is a plain-text, block-oriented format that treats a document like a tiny semantic database: narrative sections for humans, ENTITY objects for rich records, DICT/SET/ROW blocks for compact catalogs, and inline AICQL queries so AI systems can ask structured questions without leaving the file. The built-in dictionary blocks keep AI token usage low by defining aliases once and reusing them directly inside the document.

Developer Appendix Open Token Test Lab

What Chat GPT says about AI:CML ?

What Chat GPT says about AI:CML ?

1. Mission & Layers

AI:CML 1.0 keeps the “no presentation noise” spirit of the earliest drafts but remains layered enough to describe entire content sites, structured catalogs, and the queries people run on top of them—all inside one friendly, token-efficient text file. Blocks are written with UPPERCASE headers (e.g., PAGE:, ROW:). Everything until the next header belongs to the current block.

LayerPurposeKey blocks
DocumentReadable narrative equivalent to HTML articles, docs, FAQs.PAGE, META, SECTION, TITLE, P, LIST, NOTE, QA, CODE, MEDIA
EntityRich single objects (product, book, event, persona).ENTITY
Compact/tabularThousands of similar items with minimal tokens.DICT, SET, ROW
QueryInline analytics or API-like requests.QUERY (AICQL)
Design goals: 100% plain text, no XML/HTML/Markdown; dictionary-driven compression for large datasets; all semantics explicit so LLMs or classic parsers can build a JSON AST and execute queries deterministically.

2. Block Reference

Block headers are uppercase followed by a colon. Recognised headers for 1.0:

PAGE, META, SECTION, TITLE, P, LIST, ITEM, NOTE, EXAMPLE, QA, Q, A, CODE, MEDIA, ENTITY, DICT, SET, ROW, QUERY. Custom blocks may be added; consumers must ignore unknown blocks safely.

Each block stores attributes as key: value lines. Multi-line bodies (e.g., paragraph text, code) continue until the next uppercase header.

3. Document layer

Replace HTML scaffolding while keeping semantics intact. PAGE and SECTION metadata track IDs, language, ordering, and importance so downstream agents can rebuild navigation, breadcrumbs, summaries, or localization matrices.

3.1 PAGE & META

PAGE:
  id: shop-001
  lang: en
  title: Spring Catalog
  type: landing

META:
  name: audience
  value: consumer

META:
  name: region
  value: global

PAGE is required once per file. META blocks are optional but unlimited. Use them for analytics, targeting, or build metadata.

3.2 SECTION & narrative blocks

SECTION:
  id: catalog-intro
  type: intro
  importance: high
  level: 1
  order: 1

TITLE:
  Spring Catalog Overview

P:
  This catalog lists the running shoes shipping this season.
  Narrative blocks mirror HTML semantics: TITLE → h*, P → paragraph,
  LIST+ITEM → ul/ol, QA → FAQ, CODE → pre/code, MEDIA → img/video/audio.

NOTE:
  Pinned data set: products-running-2026

SECTION can reference structured data via sets:, entities:, or custom keys so readers jump directly from prose to data.

4. Entity layer

ENTITY blocks describe a single rich object without aliases. They are perfect for hero products, a featured event, or any item where readability matters more than token count. 

ENTITY:
  id: product-hero-001
  kind: product
  name: Ultra Light Runner
  category: footwear
  category.sub: running
  price.amount: 129.90
  price.currency: USD
  weight_kg: 0.72
  shipping.weight_kg: 0.90
  shipping.regions: US,CA,EU
  image.main: https://cdn.example.com/p/hero001-main.png
  image.thumbnail: https://cdn.example.com/p/hero001-thumb.png

Keys are dotted paths so downstream consumers can map them to JSON or graph structures losslessly.

5. Compact/tabular layer

DICT + SET + ROW keep large catalogs small. DICT stores aliases, SET declares which dictionary applies, ROW instances carry only alias=value pairs. This is the “table mode” that scales to thousands of rows without repeating semantic text.

DICT:
  id: dict.product.v1
  kind: product
  field.i: entity.id
  field.n: entity.name
  field.c: category
  field.sc: category.sub
  field.pa: price.amount
  field.pc: price.currency
  field.sz: size.available
  field.cl: color.options
  field.w: weight_kg
  field.sw: shipping.weight_kg
  field.sr: shipping.regions

SET:
  id: products-running-2026
  dict: dict.product.v1
  kind: product
  description: Running shoes for Spring 2026

ROW:
  i: prod-run-1001
  n: Road Runner Lite
  c: footwear
  sc: running
  pa: 89.90
  pc: USD
  sz: 40,41,42,43,44
  cl: black,blue
  w: 0.75
  sw: 0.95
  sr: US,CA,EU

Dict aliases are short (1–3 characters) so the ROW body is dense yet self-describing once the DICT is parsed.

6. Query layer (AICQL)

QUERY blocks provide embedded analytics. They reference ENTITY IDs or SET IDs, optionally specify a DICT, and expose simple line-based clauses reminiscent of SQL. Tools can evaluate these against the parsed AST to generate answers on the fly.

QUERY:
  id: q.products.over100
  from.set: products-running-2026
  using.dict: dict.product.v1
  where: pa > 100 and pc = "USD"
  select: i,n,pa,pc,sc
  sort: pa desc
  limit: 5

Clause summary: from.set or from.entity; optional using.dict; where with lightweight boolean expressions; select comma list; optional sort and limit.

7. HTML coverage & mapping

Every core HTML content element maps directly to an AICML block. Because blocks are agnostic to layout, the mapping is one-to-many friendly and survives aggressive extraction pipelines.

HTML element(s)AICML blockNotes
html/head/body meta tagsPAGE, METAPAGE captures document-level attributes; META mirrors <meta> tags.
h1–h6, section headingsSECTION + TITLElevel attribute ensures hierarchy.
p, span, div (textual)PMulti-line friendly; inline emphasis may use lightweight markup if needed.
ul/ol/li, dl/dt/ddLIST + ITEMLIST can express bulleted or numbered; description lists map to ITEM with nested keys.
blockquote, asideNOTE or EXAMPLEUse type: info|tip|warning for semantic hints.
pre/codeCODELanguage captured via lang:.
figure/img/video/audioMEDIAAttributes: type, src, optional alt/title.
faq accordions, details/summaryQA + Q + ASupports multiple answers per question.
data tables, cards, repeated componentsENTITY or DICT+SET+ROWENTITY for small sets; DICT/SET for thousands of records.

This mapping ensures “support for all HTML elements” relevant to semantic content. Presentation-only tags (layout divs, CSS classes) are intentionally ignored to keep the signal clean.

8. Full mini example

PAGE:
  id: shop-001
  lang: en
  title: Spring Running Catalog 2026
  type: landing

META:
  name: audience
  value: consumer

SECTION:
  id: intro
  type: intro
  importance: high
  level: 1
  order: 1

TITLE:
  Spring Running Collection

P:
  This catalog lists our running shoes for the spring season.
  Data is stored in a compact format for efficient processing.

SECTION:
  id: running-products
  type: concept
  importance: high
  level: 1
  order: 2
  sets: products-running-2026

TITLE:
  Running Shoes

P:
  The following data set contains all running shoe models, including
  prices, sizes, colors, and shipping regions.

DICT:
  id: dict.product.v1
  kind: product
  field.i: entity.id
  field.n: entity.name
  field.c: category
  field.sc: category.sub
  field.pa: price.amount
  field.pc: price.currency
  field.sz: size.available
  field.cl: color.options
  field.w: weight_kg
  field.sw: shipping.weight_kg
  field.sr: shipping.regions

SET:
  id: products-running-2026
  dict: dict.product.v1
  kind: product
  description: Running shoes for Spring 2026

ROW:
  i: prod-run-1001
  n: Road Runner Lite
  c: footwear
  sc: running
  pa: 89.90
  pc: USD
  sz: 40,41,42,43,44
  cl: black,blue
  w: 0.75
  sw: 0.95
  sr: US,CA,EU

ROW:
  i: prod-run-1002
  n: City Sprint
  c: footwear
  sc: running
  pa: 79.00
  pc: USD
  sz: 39,40,41,42,43
  cl: gray,white
  w: 0.72
  sw: 0.92
  sr: US,EU

ROW:
  i: prod-run-1003
  n: Trail Climber Pro
  c: footwear
  sc: trail
  pa: 115.50
  pc: USD
  sz: 41,42,43,44,45
  cl: green,black
  w: 0.85
  sw: 1.05
  sr: US,CA

QUERY:
  id: q.running.over100
  from.set: products-running-2026
  using.dict: dict.product.v1
  where: pa > 100 and pc = "USD"
  select: i,n,pa,pc,sc
  sort: pa desc
  limit: 5

9. Tooling & schemas

AICML 1.0 is the canonical plain-text format going forward. Legacy XML-based AICML deployments remain available for archival needs; consult the developer appendix for the JSON AST and parser patterns that power modern pipelines.

👉 Developer appendix: JSON AST + Parser Patterns

Next ideas:
  • Publish reference parser packages (npm, PyPI, NuGet, Maven) built on the shared AST.
  • Deliver HTML → AICML 1.0 extraction recipes for common CMS stacks.
  • Ship an AICQL evaluator + playground so embedded queries can run in CI or docs sites.