Markdown won
Markdown won. Period.
Whether we are talking about technical documentation, README files, static site generators (although for Sphinx reStructuredText is still a thing) or Tools for Thought like Obsidian - they all speak Markdown.
It is the least common denominator.
For AI, Markdown is the native format for both input and output, as well as for writing prompts, skills, or custom instructions in AGENTS.md.
Markdown is the best tool for what it does: quickly writing formatted text in a way that is easy to read even for those who are not well versed in its syntax.
Sure, it has its flaws - multiple flavors, and its simplicity might be limiting for more complex use cases like writing a book (Why I prefer reST to Markdown) or a scientific paper (where LaTeX or typst (OSS) is better suited for the job).
Some argue that Markdown is a disaster - but nevertheless, due to its simplicity and popularity, it will outlive us all.
So sure, there are issues - but:
File over app philosophy resonates with me very deeply.
Even if shit hits the fan, Markdown is just a simple, readable text file - easy to copy, sync, backup, or read on a 15-year-old ThinkPad or a gasoline-powered laptop.
Even at Anthropic, not everyone agrees Markdown is the end state: Markdown is dead, why Claude is pivoting to HTML - Thariq Shihipar (Claude Code team) sparked a massive debate by claiming he’s ditching Markdown for HTML, arguing the Unreasonable Effectiveness of HTML for AI agents.
Google Drive & Docs go native Markdown (2026-10-05) — the market verdict
ChanduThota (Google VP; 18.7k❤ / 1.09M views — one of the biggest tweets in this entire debate, captured + the Workspace update post): Markdown files now work natively across Google Drive and Docs — preview .md in Drive, open in Docs, edit, comment, and collaborate without converting the file into a Google Doc. His reasoning is the thesis of this note, stated from inside a Big Tech product org: “Markdown has become the common language between humans and AI agents — specs, plans, READMEs, tables, task lists, generated docs — so much of this starts as .md.” The friction removal framing: “no import/export loops, fewer broken formats, comments/collaboration where teams already work.” Rolling out to all Workspace + personal accounts.
This is the strongest possible datum for the “Markdown won” column: the last major holdout productivity suite didn’t adopt a Markdown mode — it dropped the conversion requirement, i.e., conceded that the file itself is the canonical artifact (file-over-app, at Google scale). Note the timing symmetry: the same week Thariq’s HTML pivot tweet (above) went viral, Google picked the other side of the ledger — HTML for rendered artifacts, Markdown for the durable text layer. Both can be true: that’s exactly the content/chrome split (Visualization).
The retrieval-side endorsement: LlamaIndex’s parsing default (2026-10-08)
llama_index (56❤) + the breakdown (both captured): “Markdown is all you need. (Mostly.)” — the argument from the parsing trenches: “a parser can get every word on the page right and still lose which column a number belongs to. Then your model has to guess.” Why Markdown is their default output: heading levels preserve section relationships (chunkers use them as boundaries, section title rides in each chunk), lists keep steps nested, simple tables give explicit row/column labels, ## beats “an object with a type, style, and coordinates” on tokens — and crucially: debuggability — “a developer investigating a bad answer can read the parsed text and check whether the problem started during parsing, retrieval, or generation.” The honest edges: merged-header tables → HTML (colspan/rowspan/<thead> blocks embedded inside the Markdown — the format boundary is per-block, not per-document); JSON for extraction targets and stored metadata; images + bounding boxes/page numbers kept when text conversion would omit (LiteParse visual grounding pairs elements with page locations). Their acceptance test is the keepable: “start with a document you know well and check that each value still has the right headers, units, and footnotes — those details tell you more than whether it looks tidy.” For this note’s market verdict lane: the parsing layer now votes the same way the editors (Google Docs), the agents (mdream’s 92% token cut), and the answer-me/trq212 output lane did — with the nuance that Markdown is the default layer, not the whole stack: HTML where merged cells need it, JSON where apps consume it, images where geometry is the payload.