The reviewsheet: one word, one page

When I want to know where things stand, I type one word: reviewsheet. A minute later I get a link. It opens a single page on my phone that tells me what was built since I last looked, asks me the decisions only I can make, and lists every piece of open work in priority order.

I type my answers into the page, tap "Copy all answers", and paste them back into the chat. That's the whole interaction.

It sounds trivial. It took about ten days, a lot of my frustration and four formats in one night to get right. And it has probably saved me more time than anything else I've built this autumn, because it replaced the thing that was costing me most: hunting through chat for what my agents had done and what they needed from me.

The problem it solves

I run my work with a lot of AI agents behind it. A factory of agents builds and checks things. Other sessions handle email, content and admin. All of them, at some point, need a decision from me, and all of them want to tell me what they've done.

For months that happened in chat. An agent would finish something and write a long summary. Another would ask three questions in the middle of a paragraph. A third would ask something I'd already answered the week before. I'm on parental leave, working mostly from my phone in a van, often in five-minute gaps. Long chat updates on a phone are hard to follow and easy to lose.

The failures that pushed me over the edge were small but repeated.

On 2 October I was sent an answer page with a "Copy my answers" button that did nothing. Two parts of the page had been given the same internal name, so tapping the button grabbed the wrong part and overwrote my answers with button text. I ended up saving the whole page as a PDF to keep what I'd typed.

The same day, a set of 14 questions about positioning came to me as menus of options. I asked, more than once: "What's your recommendation based on what you know about me and my work...?" I didn't want a menu. I wanted a view I could agree or disagree with.

And on 5 October a list I'd specifically asked for, every gap between my laptop and my always-on server, was written as prose in a chat reply. Nothing could produce it again. Once the conversation moved on, the deliverable was gone.

5 October: one word

That morning I asked the question that started it: "how do we systemise this so I can just say a simple word or phrase and always get the same format for feeding back on Claude mobile, on Mac or box, anywhere?"

Three things in that sentence mattered.

A simple word. I didn't want to describe what I wanted every time. One word, the same every time, from wherever I am. I chose "reviewsheet".

The same format. Not a fresh layout each time depending on which agent built it. The same sections in the same order, so my eye knows where to go.

Anywhere. I work from the Claude app on my phone, from my Mac when I have it, and from a small always-on server I call the box. The sheet had to work the same from all three. On the box it's served as a private web page only my own devices can reach. On the Mac it opens as a file.

The first version went up that day with four sections: what has happened, my decisions, the build order, and an "Anything else?" box. The same day I also asked for a separate outbox page for batches of drafted messages, because twenty drafted invitations dripped into chat one by one was exactly the mess I was trying to escape. "Why are we doing this drip fed in chat? Everything is getting lost." Each message on the outbox has a copy button, a button to open it in WhatsApp or email, and a tick to mark it sent. I still send everything myself.

6 October: always a recommendation

The next change was the one I'd recommend to anyone building something like this.

Every question on the sheet now carries a recommended answer. If I leave the box blank, I've accepted the recommendation. If I want to skip a question, I type "skip". Anything else I type is my answer. The page refuses to build if any question is missing a recommendation.

This changed how it feels to use. Most decisions an agent brings me are ones where it has a sensible view and just needs my yes. Making that the default means I can get through ten decisions in a couple of minutes, and only stop to think on the two where I disagree. It also forces the agent to do the thinking before it asks. If it can't say what it would do, it isn't ready to ask me.

Dark mode went in the same day, because I do a lot of this late at night.

6 October: the outcomes I had to chase

That evening showed a gap the sheet hadn't covered. I'd had to chase several things that had been built days earlier and then left: a website that was finished but never shown to me, a name change never applied, a cloud lane built and switched off. Each had "landed" as code. Nobody had taken the last step that made it real.

The cause was that the sheet showed tickets (individual pieces of work) but not the outcomes I actually care about: is the site live, has the message gone, has the invoice been paid. So I asked for one register of outcomes, shown at the top of every sheet, with each line cleared only on evidence that it's live, sent, signed or paid. "Landed" doesn't count. If I ever have to chase one, that's logged as a failure, not as me being impatient.

7 October: what was hidden

On 7 October I wrote down the format I wanted in one long, fairly heated sentence: what was built, what needs my decision with the recommended answer, and every single ticket listed in priority order so that nothing ever gets silently dropped.

That last part had just been proven necessary. Five tickets held 12 real questions for me, but they'd been tagged as questions for the factory to answer. The sheet only showed questions tagged for me, so I never saw them, and all five pieces of revenue work sat waiting. The fix was to show any question addressed to me whatever its tag, and to list every open ticket and every outcome on every sheet, with nothing filtered out.

The same morning the sheet asked me to apply a rule I'd already replaced two days earlier. I'd moved to a new way of ranking work, and an old rule was still written into the instructions. "We resolved this!" I said, and asked why I had to keep repeating myself. The rule was removed. The bigger fix was to check what the sheet asks against decisions I've already made, so answered questions stay answered.

The ordering itself comes from Jev, the scoring my factory uses to rank work by revenue importance and urgency. I won't go into it here, as it has its own write-up in Jev: how my factory decides what to build next. What matters for the sheet is that the order is computed, not chosen by whichever agent writes the page.

9 October: four formats in one night

The night of 8 to 9 October is when the format nearly fell apart, and then got fixed properly.

First, word limits. The idea was that every line on the sheet should read like an airline manual: short sentences, one idea each. The page got hard caps on how long each part could be, and refused any text that went over, naming the offending line and its length.

An hour later I judged those limits "unlikely to be good enough". So the next version went to one line per item, with tighter caps and a table that forced everything onto a single row.

That looked tidy on a laptop. On my phone, the table cut every title off after a few characters. One ticket read "#318 Ma..." and nothing else. Nobody had looked at it at phone width until a screenshot at the real size showed it.

So the third step was a mock-up, not a change. I said not to touch the real sheet until I'd seen a proposal. The mock-up dropped the table, gave each item a plain-English name on top and the reference underneath, and let the names wrap instead of being cut.

At about eight in the morning that became the fourth and final format, and it's what I use now:

1. What's been built. One line each, the line is the link.
2. Your decisions. Judgement calls only. Anything technical belongs to the factory or the agent that supervises it, and never comes to me.
3. Priority. Every ticket and outcome, in order, with a handle I can drag to reorder. My order is copied back with my answers.
4. Anything else. A free box.

Then the important part. The format is locked by a test with 12 cases. It fails if a section is added, renamed, reordered or removed, if a table appears, if a next-step column creeps back, or if a title gets cut off. Any agent that wants to change the sheet has to change the test, and that's a decision I'd see.

Two other lessons came out of that night. Questions that were really technical ("D1 to D3 are technical, so why are they asking for my decision?") had been landing on my sheet. And the copy button didn't work on my phone because the page wasn't on a secure connection, which phones require for rich copying. Both got fixed. Both were things I could only find by using it on the device I actually use.

Why a fixed format works

I've thought about why this matters so much more than I expected.

Part of it is attention. When the page is always the same shape, I don't spend any effort working out where to look. I scan the built list, answer the decisions, glance at the order, and I'm done.

Part of it is trust. A format that can drift lets things slip out quietly. A question gets mis-tagged, a section gets dropped "for brevity", a ticket goes missing because it didn't fit. Locking the format with a test means "nothing gets silently dropped" is something I can check, not something I hope for.

And part of it is that it moved the work to the right side. Before, I did the sorting: reading summaries, finding questions, remembering what I'd already answered. Now the agents do that before the page is built, and the page refuses to build if they haven't.

Make your own

You don't need a factory of agents to use this idea. If you work with AI assistants and find yourself scrolling back through chat to work out what's happened, try this.

1. Pick one trigger word. Use the same one every time, from every device.
2. Fix three or four sections. Mine are built, decisions, priority, anything else. Write them down and don't let them change without a reason.
3. Insist on a recommendation for every question. A blank answer accepts it. You'll get through decisions much faster, and the AI will think harder before asking.
4. Answer in one paste. Type into the page, copy everything, paste it back. One round trip, not a conversation.
5. List everything, filter nothing. Every open item, every time. The point is that nothing can disappear.
6. Check it on the device you'll use. A layout that works on a laptop can be useless on a phone.
7. Lock the format with a check. Even a simple one: does the page have these sections in this order, and nothing else?
8. Keep technical questions off it. If you wouldn't be able to answer it without looking up how something works, it isn't your decision.

A week on

I use it several times a day now. Mostly from the van, often late at night. The page is the update. The chat carries only the link.

It still isn't perfect. A rerun dropped some finished work from the built list once, and that got logged and fixed like everything else. But it's the first thing I've built where I stopped thinking about the tool and just used it.

If you're drowning in updates from your own AI setup and want to talk about a simpler way in, get in touch.