I've cloned Claude to create a playground where I can prototype the small interactions I long for every time I use it.

Two features made the cut for my first experiment. They share the same DNA which is less friction and less clutter while maintaining context.

Sidequests

This happens to me multiple times a day. I'm reading through a long response which I really like but there is one concept or term I need to ask a follow up question about. In vanilla Claude can highlight the text I am unsure about and Reply, but the response lands in the main thread. A few clarifying questions later and now I'm scrolling back up to find the useful answer that I now completely understand.

Do this a few times in a session and now you have a wall of text with the core insights buried between minor tweaks and clarification. Just a tangled mess of digressions.

Sidequests use the same highlight-and-ask gesture as Reply. The response just lands in a side panel instead of the main thread.

Highlight some text you want to ask about and the familiar floating menu appears but now with an extra button, Sidequest. Click it and the right rail opens with a focused composer, the highlighted passage already loaded as a chip. You ask your follow-up there, it streams its answer there. The main thread doesn't move.

Back in the chat, the original passage picks up a small numbered badge. Click the badge, the rail scrolls to the matching item. Click an item in the rail, the passage in the chat flashes a coral ring. Two-way link, no scrolling guesswork.

The reason any of this matters is that the main chat is the source of record. It's the place you scroll back to when you want to find the moment something clicked. The version of the explanation that finally landed. If every passing "wait, what about…" gets stitched into that thread, the signal-to-noise ratio collapses, and the conversation becomes less and less useful, or at least more cumbersome to navigate.

But the model isn't oblivious to your sidequests. They're appended to the system prompt as a labelled side-channel block: "the user has these sidequest follow-ups open in the side panel: 1. On passage X, user asked Y, you answered Z…". So the assistant has full context across the main thread and every open tangent. If you ask a follow-up in the main thread that ties back to a sidequest, the model can connect the dots. It just doesn't treat the sidequests as authoritative back-and-forth in the main message history. They inform; they don't pollute.

Pins

A pin button sits next to copy / retry / thumbs on every assistant message. Click it, the answer drops into the Pins tab in the right rail with a 3–6 word title.

The title is the part I'm proud of.

It's two-phase. The instant a pin is created, a regex-based title appears — a small helper strips filler openers ("Sure!", "Great question —"), grabs the first meaningful clause, and trims it. No network call, no waiting. The pin lands in the rail with a real title, in the same frame.

Then, in the background, a non-streaming Haiku 4.5 call generates a proper title. About a second later, the regex version is swapped for the AI one, with a small pulse cursor next to the title while it regenerates.

The same Pins tab supports double-click rename, hover-to-delete, and click-to-jump. Clicking a pin smooth-scrolls the chat to the source message and highlights it.

I'm still noodling on what to build next. Maybe a built in markdown editor, maybe a way to force Claude to only communicate in meme format.