Website change request for your coding agent: record it, paste one line

A website change request is usually a form: the page, what should change, how urgent it is, and a screenshot. If someone else runs your site, that form is still the right tool. If you change your own site with a coding agent such as Claude Code, you are the web team, and the request can be a screen recording that your agent reads. With Vidmatic you set up once (sign up, connect GitHub and your agent, install the Chrome extension), then each request is: record the change, choose Add to Agent Workstreams, and keep one line running in your agent: /loop 30m /mcp__vidmatic__auto-job-dispatcher. It files each request as a GitHub issue and builds your repo's open issues oldest first, so a fresh request waits behind your backlog. For one change right now, hand the recording to the agent you have open instead (show your coding agent the bug).

What a website change request form asks for, and what it loses

The forms at the top of Google ask for the same few things:

• the page the change is on,
• what should change,
• how urgent it is,
• a screenshot or file,
• and, for some, one change per request.

556 Digital's form on Tally asks "What do you need changed?", "Where is the change?" and "How urgent is this?", and asks you to "Submit one related change request at a time" (the form, read 2026-10-31). Iowa State Online's form takes a request description and "supporting files, e.g. images, syllabus, screencaps", then its team "will assign a tier after reviewing your submission" (form).

What typed requests lose is the picture. Refari, which builds and manages recruitment websites, says so in its guide to change requests: "For complex requests, a video explanation is crucial! Without this context, we may not proceed with your request to ensure accuracy and time efficiency."

The forms win on urgency and on a human approval step. The recording carries what you saw, and the work happens in your own repo without a hand-off.

The change request process with a coding agent, step by step

Steps 1 to 4 happen once.

1. Sign up

On vidmatic.ai, sign in with Google. Your first sign-in creates the account. The Free plan is $0 (pricing, read 2026-10-31); it caps how many workstreams you start each day, and paid plans have no workstream limit (limits).

2. Connect GitHub and pick a default repository

In Settings, under GitHub, press Connect GitHub. If the Vidmatic GitHub App is not on your account yet, press Install on GitHub first; for an organization's repo, an org admin has to install it. Pick the repositories Vidmatic may open issues in, then choose a default repository (GitHub docs). Add to Agent Workstreams only appears once this is done.

3. Connect your coding agent

In Claude Code, run this once per machine:

claude mcp add --transport http --scope user vidmatic https://api.vidmatic.ai/mcp

Then run /mcp, choose vidmatic and choose Authenticate. A browser page opens where you choose an access level. You do not need an API key (quickstart). The dispatcher also uses the GitHub CLI, so install gh and run gh auth login on the same machine.

4. Install the Vidmatic Chrome extension

Install Vidmatic from the Chrome Web Store. It lets you draw on your own page while you record.

5. Record the change on your own site

Open the page you want changed and start a recording from the extension. Talk as you go: "this headline should say the new price." When you reach the spot, pick up the pen and circle it, or write the new text by hand. Each mark is saved as an image at the moment you confirm it, and your agent also gets the sentence you were saying at that moment (recording your screen, what your agent receives).

6. Add it to Agent Workstreams

In the Vidmatic web app the button is Add to Agent Workstreams: on the What's next? step after you record there, on the recording's Agent tab, and in its menu in your Library. The extension build on the Chrome Web Store when we checked on October 31, 2026 (version 1.0.3) uses older words for the same action: switch its Handoff to agent button to Add to request queue and click it. Where it says "No watcher running", the watcher is the dispatcher from step 7.

The recording becomes a workstream in the Change request state: a recording with no issue and no repo yet. Sending it again opens the workstream it already has.

7. Paste one line into your coding agent

The line to paste is:

Run your dispatcher from one of your repo checkouts:

/loop 30m /mcp__vidmatic__auto-job-dispatcher

Vidmatic shows it in two places. In the web app, it appears under the confirmation after Add to Agent Workstreams when no dispatcher has checked in for 45 minutes. On Agent Workstreams, once your agent has made its first call to Vidmatic, the header reads Dispatcher never ran until one runs, and its Ask your agent button copies the line.

A checkout is the copy of your site's code on your computer. Open Claude Code there and paste the line. Every 30 minutes the dispatcher picks up the workstreams still in the Change request state, works out which repo each one belongs to, and files a GitHub issue there, on that same workstream. It stops after 3 requests or 8 minutes per run and leaves the rest for the next one (Agent Workstreams docs). Each run prints one status line, such as dispatched: #12 when it starts work on issue 12, or a line starting with idle when there is nothing to do, so you can tell it is alive. No slash commands in your agent? Call get_skill('auto-job-dispatcher') on the Vidmatic MCP server and follow it on a 30-minute loop.

What you are signing up for when you paste the line

• It runs on your machine, in your Claude Code session, with your gh login, so its work counts against your own Claude plan every 30 minutes while the session is open. Vidmatic's own skill calls Vidmatic "the ledger, never the executor". The Free plan also caps how many workstreams you start each day (limits).
• In the repo you run it from, it works every open issue, oldest first, one at a time, not only the requests you recorded. It skips issues labelled Human-Review, a few in-progress and blocked labels, labels starting with needs- or owner-, and issues an open pull request says it closes.
• If a request belongs to a repo you have not checked out, it clones that repo into the folder beside your checkout to file the issue there. Run it from the folder that holds your site's code.
• It merges by itself. Before an issue is picked up, the Human-Review label keeps the dispatcher off it. Once work has started, only --pause with that issue's number stops it.
• It re-runs your tests before merging. In a repo with no tests or CI, the agent's own review is the only check before the merge.

What your agent does after you paste the line

Take the headline from step 5. The dispatcher files it as one issue in your site's repo; a recording always becomes one issue, however many changes you mention in it. When that issue's turn comes, the dispatcher comments on it to claim it, adds a Claude-Workqueue-InProgress label, and starts a helper agent in the same Claude Code session. The helper runs the auto-implement-beta skill, which is labelled beta. It reads your repo's CLAUDE.md, AGENTS.md and CONTRIBUTING.md and follows only those for how it works.

The helper makes the change in its own git worktree, checks its own work, re-runs your tests and merges. If your repo has its own release process it uses that; if not, it waits its turn in the Vidmatic release queue, and if the queue cannot be reached it stops before merging. Then it waits for your deploy, checks production, and records a demo of the change, or an explainer when there is nothing on screen to show. The job counts as done when the issue is shipped and has that demo. On Agent Workstreams the card then shows the pull request and the demo video, so you see the new headline without opening your site.

Each request shows a Change request chip on Agent Workstreams until it is filed. Two states ask for you: Change request · needs you (the dispatcher could not tell which repo it is for; choose Pick a repo) and Change request · failed (retry or dismiss it). A request for a repo where no dispatcher runs gets its issue and waits there, shown as a Blocked item with a command you can copy.

Website change request: form, ticket or recording?

• If someone else runs your site, such as an agency or a university web team, use their form, one change per request, and attach a short screen recording when the change is visual. Refari's guide is a good model.
• If you change the site with Claude Code and want every request filed and built while you do something else, record it, add it to Agent Workstreams, and leave the dispatcher running in your repo.
• If the change is one sentence ("fix the typo in the footer"), choose New request on Agent Workstreams and type it. The same dispatcher files it.
• If you want one recording handled right now, in the repo you have open, skip the dispatcher and show your coding agent the bug instead.

The recording path asks for things a form does not: a GitHub repository (GitHub is the only tracker it files in), Claude Code, the gh CLI, and a session left running. The pen needs Chrome; without the extension, New Recording in Vidmatic records any screen, and you say out loud which element you mean.

If the request is for a teammate to pick up rather than your own agent, create a ticket from a video walks through that side.