This week's Swipe · September 16, 2026 · 5 min read
Claude in plain English
A Claude Code plugin that rewrites Claudish into plain English, a menu bar app that swaps accounts when you hit the rate limit, local memory for coding agents, and a skill that stops them quitting early.
Last week I signed off with “RIP my rate limits.” so this week I thought it’d be good to share a Mac app that lets you swap to a new Claude subscription when the first one hits the usage limit.
On the subject of Claude… Fable 5.1 is the best coding model but the way it speaks genuinely infuriates me. The odd language creeps into code too. Luckily there are plugins now that can rewrite or translate Claude back into plain English!
TLDR of the rest of this issue; dodge rate limits, publish what your agent built to a private link, memory that outlives the session, stop agents being lazy, and clone your fave expert into an agent skill.
In this issue
- ▪
Belt, One CLI for everything your agent does [sponsor]
- ▪
Claude Switcher, Swap Claude accounts when you hit the cap
- ▪
Spacefast, Publish what your agent just made
- ▪
Graft, Stop your agent solving the same bug twice
- ▪
OpenHistory, A private timeline of what you worked on
- ▪
unlazy, Stop your agent quitting before it's done
- ▪
claudish-to-english, Read Claude's replies in plain English
- ▪
ip-as-logo-skill, Cute mascot logos from one prompt
- ▪
mimeo, Clone an expert into a SKILL.md
Give every agent the same skills, tools, and context. [sponsor]
Why: Every agent you run has its own pile of skills, and none of them share. Belt is one CLI that puts the same skills, tools, and connectors in front of Claude Code, Cursor, and Codex, so a workflow you fixed once turns up everywhere.
Try: Install Belt with curl -fsSL https://cli.inference.sh | sh, then run belt skill use inferencesh/web-search and ask your agent to research something. Open Claude Code, Cursor, or Codex next and the same skill is already there.
Tools
Menu bar switching for Claude Code and Codex logins.
Why: Hitting the Claude cap used to mean a browser login as someone else. Emile Jouannet built Claude Switcher to keep your saved Claude Code and Codex logins in the Mac Keychain and swap the active one in a click.
Try: Run brew install --cask Symbioose/tap/claude-switcher, add your second account with claude auth login, and turn on auto-switch. The 5-hour and 7-day usage bars live in the menu bar, and when one account hits 100% the other takes over.
A live private link from Claude, ChatGPT, or your terminal.
Why: Your agent builds a deck or a landing page, then it sits in a folder because sharing means a hosting account. Spacefast, from Automattic, publishes it to a private link from the chat, then gives you six hours to claim it.
Try: Ask Claude or ChatGPT to publish it with Spacefast, or run npx spacefast publish from the folder, and share the link it hands back. It’s private by default and every republish keeps the same URL, so you can send it before it’s finished.
Local memory that recalls fixes and decisions across sessions.
Why: Session 27 hits the bug session 1 spent hours on, and the agent starts from zero. AEndrix built Graft as a local SQLite memory the agent writes to when it learns something and queries before it starts digging, with no SaaS or API key.
Try: Run brew tap AEndrix03/graft https://github.com/AEndrix03/Graft.git, brew install graft, then graft setup claudecode or graft setup codex. The agent gets a skill that searches memory before a hard task and saves what it learned after.
Local agents can search your day without screenshots.
Why: I saved this as a way to mine my week for things worth posting about. Zach Tratar built OpenHistory to turn your Mac’s foreground activity into hourly and daily summaries, no screenshots or keylogging, and give approved local agents a redacted read-only view.
Try: Download it from openhistory.sh, pick Apple’s on-device model (macOS 26) so nothing leaves the Mac, and let it run for a week. Then connect your agent to its MCP server and ask, “What did I actually finish this week that’s worth a post?”
Skills & workflows
Gates written up front, then rechecked before the agent says done.
Why: An agent that says “done” at 70% costs you the whole afternoon. Leon Lin built unlazy to make it write acceptance gates before it starts, split jobs into a Depth Tree where every leaf gets the full budget, and re-run every check before reporting.
Try: Run npx skills add Leonxlnx/unlazy, then ask for /unlazy tree 3 fix every failing test in this repo. It refuses to say done while a gate is unmet, and when it gives up on one it tells you instead of pretending.
A local model rewrites every Claudish reply before you see it.
Why: Talking to Claude has genuinely become like pulling teeth. Mike Gvozdev built this Claude Code plugin to run each reply through a local Ollama model and show you the plain English version, while Claude carries on speaking Claudish in the transcript.
Try: Run /plugin marketplace add gvzdv/claudish-to-english then /plugin install claudish-to-english@gvzdv-plugins, with Ollama running and a model pulled (the default is a 17GB Gemma, so pick one that fits your Mac). Every hook fails open, so if the model’s asleep you get normal Claude.
Strict rules for simple, rounded IP characters on a solid background.
Why: This might be my new favourite way to make brand mascots. Sida built ip-as-logo as a skill with strict rules: one silhouette from four to seven shapes, three colours, a character in a lower corner, and nothing fragile for the image model to fumble.
Try: Run npx skills@latest add s1dashu/ip-as-logo-skill in Codex or another agent with a top-tier image model, then ask, “Create a very simple, cute rounded ghost IP character on a solid deep navy background.” You get three directions, then six candidates to pick from.
Type a name, get a skill built from everything they published.
Why: Every expert worth cloning has explained how they think across hours of talks and essays you will never read. K-Dense built mimeo to find those sources, distill them, and write the SKILL.md or AGENTS.md itself, with every quote checked against the original.
Try: Clone the repo, run uv sync, add OpenRouter and Parallel API keys to .env, then run uv run mimeo "Naval Ravikant". Drop the SKILL.md it writes into your skills folder and ask the agent to review your pricing page the way he would.
Reach out
- Found a tool we should see? Send it over.
- Launching something? Sponsor an issue.
- Built something useful with AI? Tell us about it.
Plain .mdyou can paste straight into Claude, ChatGPT, or any agent.
Disclosure: Our picks are editorially independent. Each one has to give you something useful to swipe. Sponsors are labelled and never reviewed for pay.