Markleft

What it is in one line

A "suggestion mode" for Markdown: you annotate the document in place — text, code, tables, diagrams, SVGs — and the AI appends suggestions instead of rewriting the whole file, leaving each one for you to accept.

Who built it

Martin R. Lysk (HN: mlysk), an individual developer. A Chrome bookmarklet, an annotation spec, and an 11-minute blog post (2026-08-09); code at github.com/martin-lysk/markleft.

Read: this is a methodology-before-product project — the article first argues why the current flow is absurd, then gives you the tool. That order is itself noteworthy: it assumes readers are people who want to improve the AI writing-iteration loop, not people who want to install a plugin.

What it actually does

• Annotate in place → comment on selected text, code lines, tables, Mermaid diagrams, SVGs; image annotations use normalized coordinates (image-X-Y)
• Comments are Markdown footnotes → annotations are encoded into footnote ids (range-prev-12-chars, image-X-Y, code-line-L-col-C-len-N); a normal renderer like GitHub just sees footnotes, so the Markdown stays intact
• Stable block ids → an HTML comment markleft:block id=... before each document block makes structural changes addressable
• Append-only suggestions → suggestions are unreferenced footnotes ([^suggestion-s2-update-block-xxx]); the AI can append proposals but cannot alter the document it reviews
• Suggestion rendering in the editor → diffs within rendered elements (list items and table cells are paired before comparing); image-only replacements become before/after sliders
• Bookmarklet workflow → open a local Markdown file in Chrome, click the bookmark, get an editor with annotations and accept/reject; saving compiles a prompt teaching the AI the Markleft syntax

What it deliberately does not do: does not change the Markdown format, no cloud service, no default path of "AI rewrites the whole file."

What old behavior it replaces

The default flow of iterating a long document with AI: you describe in chat, in prose, what's wrong → AI rewrites everything → you play spot-the-difference between two versions → and you have to remember which of your original words maps to which change. The author's analogy: mailing a colleague a five-page document, getting back a prose essay of dislikes, then a fully rewritten version, and being asked to compare the two yourself to judge whether they understood — a flow document collaboration solved decades ago with comments and tracked changes.

markleft replaces the absurd chain of "prose feedback + full rewrite + manual diff" with "in-place annotation + append-only suggestions + diffs anchored to comments." The idea transfers to any AI-generated-document review scenario, not just Markdown.

Business model

None. Free bookmarklet, open source (github.com/martin-lysk/markleft), no pricing page.

Read: this is a technical/methodological essay with no trace of commercial intent. Its value is in the method, not the product form.

Hard numbers

• HN: 8 points, 1 comment (2026-08-13), Show HN; the only comment is the author's own repo link
• Blog post: 11-minute read, published 2026-08-09
• Annotatable targets: text, code lines, tables, Mermaid, images (SVG), blocks
• Users, usage: not disclosed

Four-way read

The call

A textbook case of retranslating a problem document collaboration solved two decades ago for the AI era.

Comments and tracked changes solved "how humans edit one document" in Word/Google Docs, but when AI arrived, everyone regressed to the primitive flow of describe + rewrite + manual diff. markleft doesn't invent a new interaction; it translates the constraints of an old one into a format AI can consume — append-only, addressable, acceptable or rejectable.

The transferable rule: for feedback channels to AI tools, prefer "append" over "rewrite." Appending is both a permission boundary and an auditability boundary: the AI may propose anything, but it cannot touch your original. That single constraint solves two problems at once — preventing overreach, and letting a human quickly confirm "what changed, and which comment does it answer."

Its ceiling is equally honest: a bookmarklet is a personal-tool shape, and 8 points means almost nobody saw it. The real destiny of this method is probably absorption by an editor, Claude's artifacts, or a document tool — not growing into a product itself.

What to watch next

① Whether any editor or writing tool absorbs the "append-only suggestions" interaction — validation that the method is a public good ② Whether the repo's stars/issues move — the most common fate of bookmarklet projects is author self-use ③ Whether the author grows the spec into a cross-tool standard — "comments as footnotes" is meaningless confined to one bookmarklet

What you can take from it

Product logic: any review loop in an AI content tool should copy the "append-only suggestions + addressable annotations" pair — it gives you both a permission boundary and auditability. Implementation-wise, encoding annotations into footnote ids to keep format compatibility is a low-cost high-return move.

Pricing structure: none. Free and open source.

Verdict

The method is worth something; the product hasn't grown yet. "Append-only suggestions" is one of the few designs in AI document collaboration that solves permission and auditability at once; 8 points doesn't change the value of the method. But the bookmarklet form and zero commercialization signs likely keep it at the personal-tool level. Note the interaction pattern and wait for it to be absorbed elsewhere.