This week's Swipe · August 26, 2026 · 4 min read
an agent where your mouse is
Type @@ in any Mac text field and get the answer where you already are.
You leave the Google Doc, open a terminal, paste the context into your agent, wait, then paste the answer back. Or you type ”@@” in the field your cursor is already in, and AtAt runs the agent right there.
This week: get the answer in the field you are already in, pack files and clipboard into one prompt, and publish a folder as a live link.
Let’s get to it!
In this issue
- ▪
Keep, Save anything as Markdown [sponsor]
- ▪
Zuse, Agents on real worktrees
- ▪
AtAt, Agent CLI in any field
- ▪
Relay, Pack the brief once
- ▪
here.now, Publish a folder live
- ▪
last30days, Research across Reddit and X
- ▪
caveman, Cut tokens with short speech
- ▪
asb-skills, Jason Cohen frameworks as skills
- ▪
Conveyor, Intent paperwork before code
Save ideas, notes, research, videos as Markdown for agents. [sponsor]
Why: Keep everything you read, watch, or think as Markdown for your agents. Agents can search it to deep-dive competitors, write blog posts (or newsletters like this one), track ideas and thoughts or find and summarize that article you didn’t have time to read.
Try: Capture with the Chrome extension, use the MCP to chat with your library or let agents work on their own with the CLI.
Tools
Run each agent on its own git worktree.
Why: Two agents in one directory will overwrite each other’s work. Swaraj Bachu built Zuse so each task gets its own git worktree, with the diffs, terminal, and PR context in one Mac workspace, using the CLIs you already pay for.
Try: Install the Mac app and open a mid-feature repo. Run two tasks in parallel, then check both diffs before you merge either.
Type ”@@” in any Mac text field and your agent answers right there.
Why: You leave the doc, open a terminal, and paste in context the agent could have seen from the screen. Type ”@@” in any Mac text field and AtAt runs the CLI you already have, screen context included, then drops the answer back where you were typing.
Try: Install AtAt, select a messy paragraph in a Google Doc, type ”@@”, and ask for a rewrite. The answer lands in the doc, not in a separate chat window.
Compose one Markdown prompt you can paste into Claude Code.
Why: You dictate the bug, copy the error, and grab a folder, then retype all of it into chat. Matt Sellers built Relay, a Mac menu bar app that turns voice dictation, clipboard captures, and files or folders into one Markdown prompt you paste into Claude Code, or pull over MCP.
Try: Install the DMG from the GitHub Releases page, capture one voice note and one file into the stack, then paste the prompt into Claude Code.
Tell your agent to publish a folder and get a live URL.
Why: Sharing a folder used to mean a hosting account, a CLI login, and a dashboard you never wanted. Ask your agent to publish and here.now returns a live {slug}.here.now URL, including anonymous sites that expire in a day.
Try: Run npx skills add heredotnow/skill --skill here-now -g, then tell your agent to publish a static folder you already have.
Skills & workflows
Search Reddit, X, YouTube, and HN in one pass.
Why: Google a product and you get last year’s LinkedIn posts. Matt Van Horn built last30days to search Reddit, X, YouTube, HN, and more in parallel, score by real engagement, and return one brief of what people actually said this month.
Try: Run npx skills add mvanhorn/last30days-skill -g, then /last30days the person or product you are meeting about this week.
Short answers, exact code, fewer tokens.
Why: Half of every agent reply is padding around a one-liner, and you pay for all of it. Julius Brussee shipped caveman so the chat talks in short speech while code, commands, and errors stay exact.
Try: Run npx skills add JuliusBrussee/caveman and turn it on for one debugging session.
Positioning and pricing frameworks your agent can run.
Why: Jason Cohen’s essays on positioning and pricing are useful, and you will not reopen one mid-task. He published the A Smart Bear frameworks as agent skills at asmartbear/asb-skills, so your agent can run asb-positioning or asb-raise-prices while you work.
Try: Run npx skills add asmartbear/asb-skills --skill asb-positioning and give it the product page you keep rewriting.
Confirm requirements and designs before agents write code.
Why: Agents ship unread code faster than you can review it. Kidus Tiliksew built Conveyor so work starts from confirmed requirements and designs, agents build against them, and anything that disagrees with the confirmed intent raises a flag.
Try: Confirm one requirement on a small feature, then make the agent build against it.
Plain .mdyou can paste straight into Claude, ChatGPT, or any agent.