What changes in your file when you edit on the page, how your agent hears about it, how comments work in both directions, and a few things worth knowing.
The first time the page opens, inplace adds a data-id to each block and container that doesn't have one. The ids don't change how the page looks or prints, and existing ids are never renamed.
After that, the file only changes when you make a change on the page, and then only in the block you edited or the element you moved. If the document is in git, every edit shows up as a small, readable diff.
@@ garden-proposal.html @@-<h1>A garden on the roof</h1>+<h1 data-id="h1-1">A garden on the roof</h1>-<table aria-label="Estimated costs">+<table data-id="table-1" aria-label="Estimated costs"> …-<p data-id="p-4">Volunteers water in summer.</p>+<p data-id="p-4">Volunteers take turns watering in summer.</p>
While it runs, inplace prints one line for every change made on the page, showing only the part that changed. The skill has your agent start inplace in the background and read these lines before it handles a comment or touches the file.
That way it notices a changed amount and fixes the totals and charts that depend on it, or a moved section and the “see the table below” that now points the wrong way. You can also just tell it to check your edits.
When you stop typing for a few seconds, inplace sums up what you changed, and an agent that’s following along picks it up straight away. The diagram below shows how that works when your agent waits for your pauses, as Claude Code and Codex do.
16:34:18 edited p-1: "…proposes shared raised…" → "…proposes <strong>shared</strong> raised…"16:34:40 edited td-4: "€1,150" → "€1,350"16:34:52 changed p-3 from <p> to <h2>16:35:21 moved li-1 before li-3 (inside ul-1)16:36:02 added p-5 <p> after p-2: "A structural survey last autumn confirmed…"16:36:07 paused after 5 changes: edited p-1, edited td-4, changed p-3, moved li-1, added p-5
inplace wait. inplace answers it with the pause, so your agent wakes up with exactly what you changed, and an eye appears beside each block you edited.inplace wait. That tells inplace it has finished, so the eye turns into a check and fades. If you paused again in the meantime, the new request gets that pause straight away, so nothing is missed.Other ways of following your edits skip some of these steps. When Claude Code runs inplace as an MCP server, it can get each pause as a message instead of waiting for it; the page then shows the eye but not the check, because inplace can’t tell when the agent is done. An agent that only reads inplace’s output or log file gets your edits too, but inplace can’t see it, so the page shows no marks. And any agent with a headless mode, such as claude -p or codex exec, can be started by inplace itself at each pause.
For anything bigger than a quick edit, select some text, or hover over a table, chart or row, and leave a comment on the page. Comments are kept outside your document, so they never end up in the HTML, and they survive a restart.
Your agent gets each one with the element’s data-id and the words you selected, so it finds the exact spot even after the wording has changed. Once it has dealt with a comment, the comment is marked done, with a note if there’s something you should know. If your agent’s own browser has a comment tool, as Claude’s and Codex’s do, you can use that too.
Your agent comments as well. When it notices something that’s yours to decide, such as a figure that doesn’t match your project’s data or a claim its sources don’t back up, it leaves a blue comment instead of changing your text. Answer it to tell the agent what to do, or dismiss it if nothing needs to change.
16:41:02 comment #3 on p-7 at "raised beds are shared": "Say how many beds each floor gets."16:41:40 comment #3 marked done by the agent: "Two beds per floor, as in the survey."16:44:12 comment #4 by the agent on td-2 at "€3,600": "The quote in quotes/beds.pdf is €3,840. Update it and the total?"16:45:30 comment #4 answered by the user: "Yes, please."
Press ⌘Z (Ctrl+Z on Windows and Linux) to take back a move, an added or removed paragraph, or a heading change. Each block your agent changed has a blue bar beside it until you keep or undo the change; undoing it puts back what was there before, and your agent is told you didn’t want it. To read the whole document with every change in place, turn on “Show what changed” in the list of changes: your agent’s edits are marked like tracked changes, and editing pauses until you’re done.
Bold, italics, code and links are written as <strong>, <em>, <code> and <a>, so the document’s own styles decide how they look.
Edits to different paragraphs never get in each other’s way, and if you and your agent change different words of the same paragraph, both changes are kept. If you change the same words, the document keeps your agent’s version and yours waits under “edits set aside” at the bottom of the page. There you can put yours back, add it below, or copy it. Your agent is told about it too and can combine the two for you.
On the page you can type, format, add and remove paragraphs, list items and rows, and reorder things among their neighbours. New sections or tables, moving something into another section, styling and charts are for a comment on the page.
Add data-locked to an element, such as a figure calculated from data, and it can’t be typed into or moved on the page. You can still comment on it.
Only your computer can open the document at first. Ask your agent to share it on your network and it gives you an address that works on your tablet or on a colleague’s laptop on the same Wi-Fi. There’s no login, so anyone on that network with the address can edit it. Ask your agent to keep it to this computer again when you’re done.