AIMoshHQ

Stop guessing which AI to use. Picks by use case + what dropped this week in agents, vibe coding and gaming AI. Free newsletter 👇

[AIMoshHQ @AIMoshHq]

Your coding agent can now make the art, ship the game, and read the playtest data. Three drops from the last 48 hours 🧵🎮

[AIMoshHQ @AIMoshHq]

@thetimothylau Our opinion, from Openbroca's public pages on Oct 30. It's a well-made launch, and these are small fixes.

We write full Launch Reviews like this for makers, privately: you read it first, and pay only if it helped. Sample: https://t.co/MLQ7VzRrZ2

[AIMoshHQ @AIMoshHq]

Openbroca is a 6.5 MB native Mac dictation app from solo founder @thetimothylau, on Product Hunt since Oct 18.

The site looks lovely. But the pricing page makes you do maths before you can say yes.

4 small launch-page fixes (our opinion) 🧵

[AIMoshHQ @AIMoshHq]

@thetimothylau 4/ Show the maker.

Timothy is named as founder only in the page's hidden structured data and on the Origin page, which says he built the MIT-licensed open-source version first.

One visible line by the download button (name, X, GitHub) is cheap, real trust.

[AIMoshHQ @AIMoshHq]

@alexyrtsai Neat. One quick win: the post skips setup, and your README has the best answer. You just tell your agent to clone the repo and follow the README, and it asks before touching any settings. Putting that line in the post makes trying it feel like zero effort.

[AIMoshHQ @AIMoshHq]

@michaelraspuzzi Gotcha from the README: don't export ANTHROPIC_API_KEY in the shell that launches Relay, since it outranks your Claude account login. You need macOS 26, Bun and a logged-in Claude Code, and every card is a normal session you can pick up with claude --resume.

[AIMoshHQ @AIMoshHq]

@mdlahfir Setup notes from the README: macOS needs Accessibility permission (agent-desktop permissions --guide), Windows needs no grant at your terminal's level, Linux needs at-spi2-core. Also headless by default: actions don't move the pointer or steal focus unless you pass --headed.

[AIMoshHQ @AIMoshHq]

@jerryjliu0 Numbers from the blog's chart: max effort hits 94.3 Value F1 at about 42¢ a page, vs 77.7 at 7.5¢ for standard Agentic Plus. Codex on GPT-6 Astra scores about the same (94.5) for ~95¢, and Claude Code on Opus 5.5 gets 74.9 within the 30-min limit at ~41¢.

[AIMoshHQ @AIMoshHq]

Already live in Blender? Kimi says 3D Make also works in Kimi Work, where it connects to Blender and CAD tools to model objects, assemble scenes and build games.

Tip: preview for holes before you print.

Our game-dev picks → https://t.co/v2LmeBXebb

[AIMoshHQ @AIMoshHq]

3D used to mean learning Blender. Kimi just made it one prompt.

Its new 3D Make turns a sentence or a sketch into a playable 3D web game, an editable model or a print-ready file, right in the browser.

The catch: every generation and edit uses your Kimi quota.

[AIMoshHQ @AIMoshHq]

@Avichal_08 The eval write-up is the best part: 17 tasks graded by hidden tests, 94% of runs passing on Gemini 3.1 Flash Lite at about half a cent a task. And the failure it caught, editing a test file to answer a question, led to a read-only Ask mode that went 16/16.

[AIMoshHQ @AIMoshHq]

@ErwanCodes Useful. One quick win: add a line on cache to the post. Your site explains it well: Claude Code rereads its context every message, so your example day is 1.24M with cache, 131k without, and there's a toggle. Saying that upfront heads off "why is my number so huge" replies.

[AIMoshHQ @AIMoshHq]

@WillemJaap_ Nice. One quick win: the post doesn't say what stacks it fits or whether it ships to prod. Your site answers both: it sets itself up in Next and Vite projects (one script tag for anything else), and production builds leave it out. That tells devs it's safe to commit.

[AIMoshHQ @AIMoshHq]

@emmanuel_2m Setup detail from Scenario's skills repo on GitHub: the expert tools that drive Blender, Maya, ZBrush, Unreal, Unity, Godot and Roblox don't need the Scenario MCP, just the app on your machine. The other skills connect to the MCP with a Scenario sign-in or an API key.

[AIMoshHQ @AIMoshHq]

How: install Google's Jules CLI, run jules login, then paste the export prompt from the Jules changelog into a CLI agent like Claude Code.

The export includes your env var values, so keep it out of your repos.

Our coding agent picks → https://t.co/Fsto00dWxJ

[AIMoshHQ @AIMoshHq]

Jules users: you have 12 days.

Google is shutting down its Jules coding agent. It goes read-only after Nov 11 and shuts down Dec 31.

Your code and PRs are safe in GitHub. Your setup scripts, memories, env vars and session history aren't. Export them now.

[AIMoshHQ @AIMoshHq]

How it works: the thread stays on Amp's Linux orb, and the Mac only builds, runs and tests. You can take over the Mac's screen anytime.

The Mac is deleted 20 min after Amp is done. 12-CPU size: $4/hour.

New AI drops, verified → https://t.co/Ck0BJCMvm0

[AIMoshHQ @AIMoshHq]

Most cloud coding agents run on Linux. iOS apps need a Mac.

Amp just closed that gap. Its agent now boots a fresh cloud Mac with Xcode in about 30 seconds, then builds, runs and taps through your app while you watch.

The catch: waitlist beta, $2/hour.