ICM workspaces now run inside any browser chatbot (Real Estate video example below)

Sometimes when I'm on the go, I don't want to spin up Cursor or Claude on mobile because I don't have the time to think through the prompt or my workspace, and I don't want to waste the tokens on my mobile device.

But something I'm constantly doing is telling Gemini (in Google search), or Haiku (in duck.ai) to not:

Lie
Ground its answer 100% on facts
No abbreviations
No Jargon

... and so on. You get the point, and I started thinking it would be really nice if I could use a workspace in the browser, but then I hit two road blocks:

The browser only accepts files, not folders. ICM is files and folders, so there's nothing to upload.
Even if I get it in there, how do I get the chatbot to actually walk the workspace instead of just reading it?

Both are solvable:

Roadblock 1: one file, not a folder.

I was hoping the browser agents would accept a markdown file, but that didn't work and the ai model would respond with a timeout. Then I tried a text file, same problem. Then I noticed that models would accept PDF as a format. So then, I flatten the whole workspace into a single file, stamping each file's path at the top of its contents so the model can still tell the stages apart. Nothing is lost. The router, the stage contracts, the references are all there, in order, in one upload.

Roadblock 2: getting the chatbot to run the workspace.

I put one instruction at the very top of the flattened file: when my message starts with >, treat everything after it as the request and walk the workspace on it. Then I upload the file, type > and whatever I need, and the chatbot runs the pipeline instead of describing it.

So now, on my phone, from a plain browser tab:

> how far is the moon from earth

comes back ground-in-facts, no jargon, no abbreviations, every rule I wrote once, applied automatically.

No Cursor, no app, no wasted tokens.

So, at first I wrote a workspace that prevents drift and I don't have to repeat myself. Then I got to thinking, this would be nice for typing in a city, and getting the latest news, city stats, real estate postings etc by using this, and this idea would really stress test if this concept works on ai search.

So I built a second workspace: type in a city, get back filtered listing links (Zillow, Redfin, Trulia, Realtor), area and city stats with sources, and the latest local news, all in one report I can save. On my phone:

> homes under 300k in Austin, TX

comes back as a clean report, sourced, with clickable links. That one request leans on everything a browser AI can do, searching, sourcing, and linking, so if the concept held up here, it held up anywhere.

It held up. But the surprise was this: the same file runs differently depending on which chatbot opens it.

Three things made the difference between a report that works and one that doesn't.

1. Some chatbots can look things up while they answer; some can't, and that changes what you get back.

Gemini (inside Google Search) can search the live web as it writes, so it returns current listings, real statistics, and today's news with the sources attached. Haiku (in duck.ai) can only use what it already learned during training, so it has no way to look anything up mid-answer. Instead of guessing a number it can't confirm, Haiku labels it "unverified." I write each stage so it produces a complete, clean report either way: Gemini fills in live data, and Haiku fills in what it knows and flags the rest.

2. A link is only clickable if it points to a real page that already exists online.

When Gemini writes a link, it first checks that the web address leads to a specific page in Google's search results, like one actual home listing or one published article. If the address checks out, the link works. If I instead hand it a "search Zillow for Austin" address, that address doesn't lead to one fixed page, so Gemini can't confirm anything is there and prints the address as gray text you can't click. So for chatbots that can search, I give direct links to real pages. For chatbots that can't search, I give the search links, and those tools show them as clickable anyway.

3. The note at the top of the file has to ask, not shout.

The first thing the chatbot reads is a short note telling it what to do with the file. When that note is written as hard commands ("YOU MUST DO THIS"), Gemini decides the file is a document to summarize, and it reads the file back to me with citations instead of running it. When the note is plain, like "if my message starts with >, run the workspace on what comes after it", Gemini treats it as a job and does the job. So I keep that top note calm and put all the strict rules inside the workspace, where the chatbot follows them instead of quoting them.

The one catch: a chatbot can still ignore a rule you wrote. Gemini once used a source I had specifically told it to avoid. The workspace sets the intent, and the chatbot makes the final call, so I read the finished report to confirm it, rather than trusting the instructions alone.

Here are two videos using the same flat pdf file that I turned into from a workspace. (I ran out of attempts on duck.ai but the results are better because of the google dorking methods that haiku can leverage and gemini strips out).

If you're interested. I can post the two flat pdf files for others to use too. But I felt this is a big win for ICM (with limitations of course) and wanted to share!

0:38

0:24

10 comments

ICM workspaces now run inside any browser chatbot (Real Estate video example below)