reviewbotGitHub

A Claude Code plugin

Your agent builds. You decide. One pass.

An agent can make 60 things in an afternoon, and then it asks you about all 60 in chat, and the questions scroll away. The building part is fast. The bottleneck is you.

With reviewbot you kick off a huge session and go to sleep. When you wake up, or when you hit the cafe for a spro, the whole batch is waiting on one page. You review it from wherever you are.

65 seconds. Claude Code drew every frame in code, and it was reviewed on a reviewbot page: two rounds, music picked by ear, done.

Install

Two lines in your terminal

$ claude plugin marketplace add hntw/reviewbot $ claude plugin install reviewbot@reviewbot

Needs Claude Code with Artifacts publishing (private pages on claude.ai) and Python 3. Then ask your agent to “put this on a review page,” or just say “reviewbot.”

How it works

The loop

  1. Build somewhere safeThe agent does the whole batch on a test site, a draft or a branch. Nothing live.
  2. One pageEvery decision lands on one private Claude Artifact, with real screenshots and one question per card.
  3. You tap through itApprove, Change or Hold, plus notes. Answers save as you tap, on any device.
  4. The agent reads it allChanges first, then holds, then approvals. Every note, approvals included.
  5. Round 2A new page quotes what you said and shows what changed.

How it actually works

All inside Claude

Claude Code does the work

It runs on your machine. It does the batch, takes screenshots, and builds the review page.

The page is a Claude Artifact

A private page on claude.ai that only you can open unless you share it. It has its own little database, so every tap and note saves the moment you make it. That's why it works on your phone, your laptop, the Claude app or a browser, and why you can stop halfway and pick it up later. (Yeah, we didn't know Claude could do that either.)

“Done” is all you say

Claude Code reads your answers straight from the page's database. No exporting, no copy and paste. It sorts them (changes first, then holds, then approvals), reads every note, and gets to work.

The ship rule lives in chat

Anyone you share the page with could tick a box, so a box isn't proof it was you. Your own words in the Claude Code chat are. That's what lets it push to the live site.

It gets smarter

Lessons go in a plain text file on your machine, and the agent reads it before every review. The next project starts where the last one left off.

The rules it follows

Built so nothing ships by accident

Approve isn't ship

A tap says “this is right.” Anything that goes live waits for your OK in chat. The page hands you the exact line to paste.

Every note gets read

“Love it. Also do X.” is a new task, not a done one.

One question, one decision

Approve always means one thing. Two decisions get two questions.

No fake done

If something half worked, the next page says so, with the fix.

What's inside

One skill, five small files

The skill teaches your agent the loop and the rules. The kit builds the page and reads the answers back. About 90 tokens sit in every session; the rest loads only when you use it.

Your agent keeps its lessons in ~/.claude/reviewbot/LEARNINGS.md, outside the plugin, so updates never wipe what it has learned.

skills/reviewbot/
├── SKILL.md            # the loop and the rules
└── kit/
    ├── make.py         # review.json -> the page
    ├── kit.css         # the look
    ├── kit.js          # buttons, notes, saving
    ├── answers.py      # answers -> a table
    └── review.example.json