← Blog

HTML Is the New Markdown. Reviewing It Is the Missing Half.

By Marcin, founder of Markloop · July 2026

In May 2026, Thariq Shihipar from Anthropic's Claude Code team published a post arguing something that sounds cosmetic and isn't: agents should stop writing Markdown and start writing HTML. It went viral - follow-ups on Lenny's Newsletter and InfoQ, and a wave of people converting their planning workflows overnight. One tester ran his 20 examples head-to-head and scored HTML the winner in 17 of them.

The thesis, in one paragraph

Models are brilliant at Markdown. That was never the problem. The problem is the humans in the loop: a wall of Markdown gets skimmed, an interactive HTML page gets engaged with. Give a plan color, layout, diagrams, collapsible sections, toggles - and the person reviewing it actually stays connected to the work. As Shihipar put it, HTML makes it easier "for us, the humans in the loop, to visualize, interact with, and stay connected to the work being done." Specs, implementation plans, code review walkthroughs, data exploration - anything that benefits from a tailored interface benefits from HTML.

Same content, different medium Models are brilliant at Markdown. Humans engage with HTML. plan.md ## Implementation Plan ### Phase 1 - Data layer - add orders table + index - backfill script (batch 500) - [ ] decide retention policy ### Phase 2 - Sync loop - webhook consumer w/ retries - reconcile job nightly - [ ] SLA for stock holds? ### Risks - duplicate events on replay - ERP rate limits unknown gets skimmed implementation-plan.html Implementation Plan Phase 1 · done ✓ Phase 2 · in progress data layer sync loop rollout ? Open question: SLA for stock holds? click to answer - right here in the doc ▸ Risks & mitigations (collapsed - expand) ▸ Rollback plan gets engaged with
Same content, different medium - a Markdown wall gets skimmed; an HTML page gets engaged with.

We agree - we bet the company on it

Markloop exists because of the exact dynamic Thariq describes, seen from the other side of the desk. When your agent's output graduates from Markdown notes to magazine-quality HTML documents, something quietly breaks: the collaboration layer doesn't come along.

Markdown had a home - it lived in the repo, reviewed in pull requests. Google Docs content had a home - comments live right there in the margin. The bespoke HTML artifact your agent just produced has no home. It outgrew Docs (paste it in and the layout, diagrams, and code blocks die) and it was never in the PR flow. So feedback regresses to the stone age: screenshots in Slack, "see my comments in the email below", and you - the person who didn't want to be a middleman - hand-translating every remark back into a prompt.

It's telling that Shihipar's own list of desired workflows includes "adding comments and annotations to HTML plans." The authoring half of the shift shipped. The reviewing half didn't.

What the reviewing half needs

Authoring got…Reviewing still needs…
Agents that produce rich HTML in secondsA place where anyone can open that HTML and mark it up - no tooling on their side
Artifacts: live pages at a URLComments anchored to the exact element or sentence, not "the third paragraph, I think"
Versions republished in placeVersion-aware feedback: what's addressed, what's still open
MCP: agents that can push work outThe reverse: feedback flowing back into the agent as structured context, not copy-paste

That last row is the one we care most about. A comment that says "§2 contradicts §3" is only useful if it arrives in the agent's context with its target, quote, and version attached - ready to apply. Otherwise a human retypes it, and the loop has a manual crank in the middle.

The loop, closed

Here's what the full circle looks like when both halves exist - this is Markloop's flow with Claude Code, and the pattern generalizes:

  1. The agent writes the doc in HTML - spec, plan, report - exactly as the thesis prescribes. It embeds open questions where decisions are needed.
  2. People respond in place. Reviewers - teammates or a client, no account, no terminal - comment on the exact spot and answer the embedded questions in the rendered page.
  3. The agent pulls the feedback over MCP as structured threads (target, quote, intent, version), applies the changes locally, and publishes v2. Addressed comments get checked off; open questions carry forward.

HTML made the document worth engaging with. The loop makes the engagement worth something.

The Markloop loop Add → Share → Collect → Pull → Apply ↻ your coding agent Claude Code · Codex 1 · writes the doc (HTML) 2 · markloop_put_version() 4 · markloop_pull() 5 · applies → publishes v2 ✓ order-sync.html v1 → v2 3 · People review 1 “What if the reservation times out?” answered ✓ 1 open → next round push · MCP pull · MCP feedback comes back structured - target, quote, intent, version - and it repeats Reviewers need no account and are always free · the agent applies changes locally - nothing runs on Markloop's servers
The missing half, closed: humans respond in place, the agent pulls it back structured.

Your agent already writes HTML.

Markloop is where people review it - anchored comments, answered questions, versions - and where your agent picks the feedback back up over MCP.

Start free - 14-day trial, no card required →

Reviewers are always free.

Related: What are Claude artifacts? · How to share a Claude Code artifact outside your organization