Source: raw/Claude_Code_+_Drawbridge_Builds_Websites_By_Sight_NEW_Plugin.md
Creator: Robo Nuggets (host Jay) | URL: https://www.youtube.com/watch?v=K0j_ADBkN9I | Duration: 8:51 | Platform: YouTube | Published: 2026-03-26
Repo: breschio/drawbridge | Stars: 969 (2026-10-27) | Language: JavaScript (Chrome extension)
Drawbridge is a Chrome extension that lets you click on your running app, leave a comment or draw a box, and turn each note into a task file your coding agent can work through. It fixes the gap of describing a visual problem in text. Each task carries the screenshot plus the selector, element context and bounding box, so the agent gets more than pixels. Jay uses it to polish UI from his phone, sending Claude Code “pick up my Drawbridge comments” over Telegram.
Key Takeaways
- Two annotation modes. Press
Cto comment on a specific element. PressRto draw a rectangle over an area, for layout or spacing problems not tied to one element. Notes queue in a dockable drawer (bottom, left or right). - Files the agent reads (per the repo README):
.moat/moat-tasks.md(readable list),.moat/moat-tasks-detail.json(title, comment, bounding box, screenshot path),.moat/screenshots/, plus a generated.claude/commands/bridge.mdslash command. The captions’ “mot_tasks.md” ismoat-tasks.md. - Three processing modes.
- Step: one task at a time with approval.
- Batch: related tasks grouped, one approval per group.
- YOLO: everything, no approvals.
Run
/bridge, or say “run Drawbridge in step mode”. In Jay’s demo the agent marked tasks “doing”, then “done”, and cleared the log afterwards.
- Why it beats a screenshot. “If you just take a screenshot, all that really gives your AI model is just pixels.” Drawbridge also passes the HTML and selector context, so the agent can find the right element.
- Works beyond Claude Code. The README lists Cursor, Codex and Windsurf workflows too.
- Licence caveat. Jay calls it free and open source. The README is a custom licence: free to use, copy and modify, but redistribution, selling or offering it as a commercial service needs a commercial licence. Using it on client projects appears allowed; reselling it doesn’t.
Try It
- Clone the repo and open
chrome://extensions. Turn on Developer mode, click Load unpacked, and pickdrawbridge/chrome-extension/. - Open your local app and click the Drawbridge icon, then Connect Project, and choose the project root.
- Leave three comments (
C) and one rectangle (R) on things that look off. - In Claude Code, run
/bridgein step mode for the first session to see how it reads the tasks, then batch mode once you trust it. - For
file://pages, enable “Allow access to file URLs” in the extension details.
Implementation
Tool/Service: Drawbridge Chrome extension plus the /bridge command it installs.
Setup: Unpacked Chrome extension; connect it once per project folder. Jay had Claude Code clone the repo and walk him through installation.
Cost: Free for direct use (custom licence; commercial redistribution licensed separately).
Integration notes: Drawbridge adds its generated folders to .gitignore. It pairs naturally with any build-then-polish loop, such as the cinematic sites pipeline or Claude Design handoffs to Claude Code.
Open Questions
- How Drawbridge compares with Claude in Chrome for visual QA isn’t covered. Claude in Chrome lets the agent look for itself; Drawbridge is for you pointing.
- How it behaves in single-page apps where selectors change between builds wasn’t tested in the video.
Related
- Reddit Proto — click-to-annotate edits sent to Cursor inside a design-system prototyping harness.
- Claude in Chrome — agent-driven browser inspection
- Frontend Design Deep Dive — design skills for the code that Drawbridge corrects
- Impeccable v3 — automated design audit to run before manual annotations
- Channels — the Telegram path Jay uses to trigger task pickup
- Cinematic Sites Agent — a build pipeline whose output needs this kind of polish