← Back to devlog
Tools · September 2026

Markdown Can't Show Me a Font

I wrote about workdoc back in June. An agent takes its pile of open design decisions, writes them into a markdown file as “forks”, and I answer them at my own pace in whatever order suits me. It reads the file back, writes up what I settled, asks whatever new questions fell out, and keeps going.

It fixed the exact problem I built it to fix.

Then a session asked me to pick a typeface.

It did everything right. A fork per text role, the candidates listed under each one, a description apiece, a recommendation with real reasoning attached. Textbook protocol.

Spectral. EB Garamond. Alegreya. Adjectives about them.

I have no idea what any of those look like. Neither do you.

The File Was Never the Point

The insight workdoc is built on holds up fine. A wall of decisions does not fit through a one-channel chat, so you put the wall somewhere I can walk along it. That part was right.

Markdown was just what I had lying around.

And markdown is passive. It cannot render a typeface. It cannot show me a colour. It cannot check that I answered the thing it asked, cannot show me only what’s still open, and cannot let me choose between three options it already handed me without me retyping one of them as prose under an #### Input heading. Every fork, no matter its shape, arrives as the same blank region in a text editor and leaves as a paragraph I had to compose.

For a decision like “should the resolution live in git”, a paragraph is exactly right. For “which of these nine faces looks least stupid on a title screen”, it’s absurd. I’m being asked to imagine the thing and then write an essay about the thing I imagined.

So keep the protocol. Change the paper.

That’s workpage.

How It Works

An agent publishes an interactive HTML page carrying the same four slots per fork that a workdoc has. Description, Recommendation, Input, Resolution. The change is that Input stops being an empty region and becomes a real control.

A radio group where the options enumerate, with the agent’s recommendation already selected, so my most common action is confirming a reasoned default instead of retyping it. A text area where prose is genuinely what’s wanted. Rendered type specimens, actual swatches, tables and comparisons drawn instead of described.

My input saves server side. I work the page whenever, in whatever order, leaving half of it open for a day if I feel like it. When I’m ready I ping the agent in chat, same as before. It reads back every fork’s state at once, writes up everything newly answered, opens whatever new questions fell out, and republishes.

Same protocol. The paper finally matches the decision.

Have a Look

Describing an interface in prose has the same problem as describing a typeface in prose, which is a joke I am apparently not finished making.

So here is one you can open. The workpage that designed workpage carries the six forks from the next section as a real page, instead of the workdoc they were genuinely settled in. Radio groups with the recommendation already marked, a text box on every fork whether it looks like it needs one or not, resolved forks collapsed with their full text still under them.

Nothing saves, nothing phones home, and it renders fine with your network off. That last part is load bearing, and the section below explains why I had to learn it the hard way.

The Calls That Mattered

Six forks went into designing this, settled across two passes. In a workdoc, obviously.

The page is the record, not a rendering of one. My agent recommended the opposite. It wanted the page to be a pretty input device that writes resolutions back into markdown, so the decision still lived in git where every other decision here lives. Reasonable, and it was arguing against a mistake this workspace has actually made before. I overrode it. Archive the page itself into the project’s artifacts folder and reference it from the journal, which turned out to be a convention I already run for everything else. That folder is in git. 202 files of it. The durability objection answered itself once somebody bothered to check.

There is always a text box, on every fork, no exceptions. This is the highest value rule in the whole spec and it reads like filler, which is why it’s pinned to the top of the skill where an editor can’t quietly trim it. On the first real run I overrode the recommendation on two of four forks, and on both of them my actual reason arrived through the free text box rather than the choice control. One of them was a bigger claim than the fork had even asked about, and became a new fork. A choice-only control would have thrown away both. The caveat attached to a pick is regularly worth more than the pick.

Server side store, not local storage. The pattern I saw this idea in used browser local storage, which is sandboxed per origin and per profile and cannot be read by a terminal agent at all. Every local-storage version of this needs a bridge. Export a file, run a little server, drive a browser and evaluate in page context. That friction is what kills the pattern in practice, so it’s worth being precise about instead of substituting silently.

A chat ping, and nothing built to replace it. The agent could poll. The page could fire a webhook when I save. Both are worse. A page that kicks off an agent pass while I’m still deciding is actively hostile, and more importantly a notification hook would wire the protocol to one specific agent service. The ping keeps the page a dumb portable surface. Anything that can read the store can run a pass, including some tool I haven’t installed yet.

Bake the answers into the published HTML. This one was a genuine bug in my own spec and I only caught it by asking an idle question. If the answers live only in the store and get read at runtime, then no published version of that page ever contained them, and saving the page to disk archives an empty form. Which is worse than archiving nothing, because it looks complete. The first pass of the pilot page was destroyed this way and had to be rebuilt from a transcript. The test is blunt: open the archived file from disk with the machine offline. Blanks mean the export is broken.

Archives are stamped per pass. .pass1.html, .pass2.html. A rolling filename silently eats history and doesn’t tell you.

Sixteen Forks and a Sidebar

The pilot ran six forks on one screen, and v1 specified a running tally because six forks on one screen is all a tally has to survive.

Then a session built a page with sixteen forks across three passes and added a sticky sidebar rail, unprompted, because the tally had stopped working. A tally tells you four are open. It doesn’t tell you which four, on a page far too long to see at once.

Three things about that rail are load bearing and are now in the spec. Each entry gets a short handle written by hand, distinct from the fork’s sentence-long heading. “Full bag on equip” instead of “What happens when you equip an item and the bag is full”. That’s an authoring step, not a truncation.

The rail groups by provenance rather than by topic. Carried over. From the critique. Opened by pass 2. On a multi-pass page, where a question came from is both the hardest thing to reconstruct and the best signal of how much it weighs.

And one status function drives both the rail dot and the fork’s own marker, so the summary physically cannot drift from the thing it’s summarizing.

The rail is required above roughly eight forks. That number is set from two data points, not from a principle. It’ll move when a page in between tells me otherwise.

Where It Stands

Workpage is the standard instrument now. The proof landed in two parts, and I only planned one of them. The pilot proved the medium, on the exact axis I’d argued for it. The sixteen-fork page proved the scale, which I hadn’t thought to worry about. A format that works at six and falls over at sixteen would not have been worth replacing anything with.

Workdoc is still here, demoted. Proving workpage works is a different claim from proving workdoc is unnecessary, and conflating those two is how a working fallback gets deleted on momentum. Workdoc is a file on disk with no publish step and no platform underneath it. The honest cost of keeping it is drift, so it has an expiry date on the calendar. If nothing reaches for it by October 4th, it goes.

One exception showed up almost immediately, and it’s a good one. Approving screenshots looked like a perfect fit for the format, except the page-is-the-record rule forbids recycling a page, and every screenshot approval is already a commit of a baseline image. The history that rule protects was being kept by git the whole time. So a workpage can be recycled when something outside the page already records the decision durably, and nowhere else. Narrow, testable, not a convenience.

I picked Spectral, by the way. Took about four seconds.

I could see it.