Human Review

What Human Review is

Human Review (github.com/petergyang/human-review) is a small JavaScript CLI, MIT-licensed, first published in late July 2026. You install it as a skill your agent can invoke, or run it directly with npx, then point it at a file — `/human-review (your file)` — and it opens that file in your browser as an editable page. The problem it names is a familiar one: giving an agent feedback in chat means writing a paragraph like "in the third section change X to Y, cut the card that repeats the first one, rewrite the call to action", then re-reading the whole file to check what the model understood.

• Edit the output directly. Change text in place with basic formatting, make bulleted and numbered lists, add or edit links with a keyboard shortcut, resize images by dragging a corner, move images, and drag whole blocks to rearrange the page.
• Comment like a document. Select a phrase and leave a comment anchored to that exact text, or click an image, chart or section to comment on the element. Deleting something needs no explanation in chat — you just remove it.
• Send it all at once. One Send hands your agent every edit and comment together, and it applies them to the source and refreshes the page for another pass. Command-clicking links lets you review several pages in one session without losing feedback along the way.
• Review a running app, not just a file. Pointing it at a localhost URL reviews the live page. The README is precise about the difference: for HTML files, direct edits and resizes save automatically, while Markdown and localhost reviews need Send so the agent can apply them back to the source.

Where it fits in a founder's stack

This is not a growth tool, and it is worth being clear about that. It sits one layer earlier, on the round trip that every AI workflow creates and almost none of them solves: the gap between the agent produced something and this is right. Describing edits in prose is lossy in both directions — you spend longer specifying a change than making it, and then you have to audit whether it landed. Editing the artifact directly removes the specification step entirely.

The fit is sharpest for marketing pages, which is where an agent's output is most likely to look finished and be wrong. If you are having a model build a landing page, the review pass is the whole difference between a draft and a site — see should I use AI to build my website for the specific faults worth hunting. The repo is young and actively developed rather than settled: created in July 2026, with v0.5.0 and v0.6.0 landing a day apart in early August and v0.6.1 on 2026-08-15. Expect the interface to keep moving.

The failure mode a review step adds

Putting a human in the loop is the right instinct, and it is not free — it introduces a new way to lose work, which is the review that never happens. We know this from our own platform, where every outbound action an agent proposes waits on a person approving it, and the gate is enforced in code rather than requested in a prompt, because a prompt-level rule is a request and a code-level one is not. Of the first two dozen approval requests our system ever produced, exactly half never became a completed action: five failed at the other end, and seven were simply never clicked.

The seven are the interesting number. Nobody rejected them; the queue just outlived the moment when anyone cared. That is the tax a review step charges, and it is why the design details matter more than the principle — the queue has to be visible, the batch has to be small enough to clear in one sitting, and the tool has to report back what actually happened. We had to add the last part after the fact, because an agent that cannot see the outcome of its own approval will cheerfully tell a customer that something shipped when it did not.

That boundary is the whole design of AgentCeres — the AI Growth Officer at agentceres.com. A roster of specialists drafts the SEO pages, social posts, ads and outreach; anything that goes out to the world waits on a person clicking approve. A local editor like Human Review makes that click a better-informed one, because reviewing the artifact beats reviewing a description of it. The draft is the cheap half either way — see what an AI marketing team can actually do.