The creator’s real bottleneck isn’t content — it’s the waiting
Every social media operator I know has the same dirty secret: the most expensive part of our workflow is not writing captions or editing Reels. It’s the dead time between “the machine is working” and “the machine needs me.” You fire off a batch of AI-assisted drafts, a repurposing pass, a thumbnail generation run — and then you tab away to check Instagram, answer a Slack ping, or “just quickly” look at analytics. Twenty minutes later the run is sitting there, finished, waiting for a human decision you forgot to make. Multiply that by every tool in your stack and you’ve found the real tax on a one-person content operation.
That’s why CrewTower, a small macOS utility that lives in the MacBook notch, caught my attention this week — not because I’m the target user (it’s built for coding agents like Claude Code, Codex, and Cursor), but because the workflow problem it solves is structurally identical to the one every social team has with AI content pipelines. The maker, Said Altan, shipped a new Claude Code “mod” integration that lets you fire the next instruction into a running session from the notch — no terminal-tab hunting. If you run any kind of AI-assisted production line, the design lesson here is worth more than the app itself.
What CrewTower actually is (and what it isn’t)
Let me be precise, because the Product Hunt page is easy to misread. CrewTower is a Mac-only app that sits in the notch and shows every active agent session at a glance: which are working, which are waiting on your answer, which are done. At launch in August it could watch sessions and answer permission prompts. The new version, 1.0.33, adds a Claude Code mod — a small JavaScript or TypeScript file that Claude Code loads into its own process and calls on every event — so a card in the notch can now take Claude’s next prompt, stop a running turn, compact the conversation, and show the same context breakdown that /context draws.
The pricing is refreshingly boring: $9.99 once, one Mac, no subscription (per the maker’s launch post). It requires Claude Code 2.1.287 or newer and macOS 15. Approvals still cover Claude Code, Codex, Cursor, Gemini CLI, Qwen Code, and opencode. The mod talks only to the CrewTower app over a local socket — nothing goes to a server, it stays off until you enable it, and disabling it uninstalls it. The maker is explicit that the mod is still in beta.
Here’s the part creators should internalize: this is not a social media tool. There is no scheduling, no publishing, no analytics, no asset library. If you came here looking for a Buffer or Later replacement, close the tab. What CrewTower is, is a supervision layer for long-running AI work — and that’s a category social teams are about to need whether they realize it or not.
Why TikTok creators should care more than LinkedIn ones
My take: the operators who will feel this pattern first are short-form video people, not B2B writers. A LinkedIn ghostwriter’s AI workflow is mostly text generation — fast, cheap, low-stakes to interrupt. A TikTok or YouTube Shorts pipeline is a chain of slow, expensive, stateful jobs: script generation, voiceover synthesis, CapCut template assembly, thumbnail variants, caption localization. Each step can run for minutes, each step depends on the last, and each step can silently stall waiting for a human “yes.” That’s exactly the failure mode CrewTower was built to fix for code. The video people just haven’t been given the equivalent yet.
The real problem: agents stall, and stalled agents burn money
The maker’s own framing is the most honest thing on the page. In an earlier post he describes the problem plainly: he’d be doing other work or scrolling social media, and his sessions would “just sit there waiting for input,” wasting time and tokens. That’s not a coding-specific complaint. It’s the economics of any metered AI workflow.
When I scheduled a month of content across five platforms recently, the AI-assisted parts of that batch — caption variants, hook rewrites, alt-text generation — were cheap per call but brutal in aggregate because I kept context-switching away mid-run. Every abandoned session is either wasted tokens (you paid for the generation) or wasted wall-clock time (you’re paying a human to wait). The notch solves the attention half of that equation by putting session state at eye level, “no matter what app you’re in,” as the maker puts it.
Where the math breaks
I want to be careful here, because the Product Hunt page makes no performance claims and I’m not going to invent any. But the structural argument is sound: if your AI pipeline has N steps and each step has a p% chance of needing human input, your expected idle time scales with N × p. Content repurposing is the poster child — one long-form video fans out into a Short, three Reels, a carousel, a thread, and a newsletter blurb. That’s six or seven AI-touched artifacts per source asset. Supervising them from a terminal is a full-time job. Supervising them from a persistent glanceable surface is a background task. That’s the whole pitch, and it’s a good one.
What social teams can borrow from this design
Even if you never install CrewTower, three ideas in it translate directly to how you should be running AI content ops.
1. Make waiting states visible, not discoverable. The notch is a genius piece of real estate because it’s always in peripheral vision. In your own stack, the equivalent is a single dashboard or even a pinned note that shows every in-flight AI job and its state. If you’re running batch generation through Zapier or Make, pipe completion events into one Slack channel. The point isn’t the tool; it’s that “waiting on you” should never require you to go looking.
2. Queue the next instruction before the current one finishes. The mod lets you type Claude’s next prompt and press Send; if Claude is still working, it runs when the turn ends. That’s a tiny feature with a big operational implication: you can plan ahead instead of reacting. In content terms, this is the difference between approving captions one at a time and pre-loading the next three review decisions while the render finishes. The maker confirmed in the thread that you can’t yet edit or cancel a queued message — more on that below.
3. Watch context, not just output. The card shows the same context breakdown /context draws and hints when to compact. For creators, the analogue is watching your prompt context — brand voice docs, past examples, platform specs — and knowing when it’s gone stale or bloated. A bloated context is why your AI captions start sounding like everyone else’s. Compacting at a task break is good hygiene; doing it mid-task is how you get tonal whiplash.
The “pick a spec file” idea is the one to steal
In the comments, Leo Heo asks for the ability to pick a feature spec file and send it as the next prompt instead of retyping context every time. The maker says it’s not built yet but he’s thinking about it. This is the single most valuable idea on the page for social teams. Your equivalent of a spec file is a brand-voice brief, a platform style guide, or a reusable prompt template for “turn this transcript into a 30-second hook.” If your AI tooling makes you re-paste that context on every run, you’re burning tokens and introducing drift. Build the template library now, regardless of what tool you use.
Where my judgment says it falls short
Platform lock-in is severe. Mac-only, macOS 15 minimum, Claude Code 2.1.287 minimum, and the mod is Claude-specific. A commenter named Kirtik asks for a Windows version; the maker doesn’t address it. If your team is on Windows or older macOS, this is a non-starter today. Not disclosed: any roadmap for other platforms.
The mod is beta, and the maker says so. He explicitly asks users to “tell me what breaks.” Treat any production dependency on it accordingly.
No edit/cancel on queued messages. Denys Vysochin asked whether you can edit or cancel a message while Claude is still working; the maker’s answer is no — once sent, it waits and goes in when the turn ends, and you can only hit Stop on the card. He says he’ll “probably add that.” For content workflows where the AI’s output changes what you want to ask next, that’s a real friction point.
The “task break” detection is naive. Gene Dai asked how the card decides it’s at a task break — turn ending, or something smarter like tests passing or a commit? The maker admits it’s “pretty simple”: it fires when the turn ends and Claude is waiting, and only suggests compacting once context gets big, around 300k on 1M. He concedes that tests passing or a commit “would be a much better moment.” For creators, the smarter trigger would be “asset exported” or “draft approved” — the tooling isn’t there yet, and that’s an honest gap.
This is not for you if you’re a social-only operator with no coding-agent habit, you’re on Windows, or you want a tool that touches publishing at all. It’s also not for teams that need shared/multi-user state — nothing on the page suggests collaboration features, and at $9.99 one-Mac it’s clearly a solo-operator product.
What I’d watch / test next
This week, do the low-tech version of what CrewTower does before you spend a dollar. First, inventory your AI content pipeline and count how many steps can silently stall waiting on you — that number is your idle-time tax. Second, consolidate every “needs human” signal into one always-visible surface (a pinned Slack channel, a persistent dashboard tab, whatever you’ll actually glance at). Third, write down your three most-repeated prompt contexts as reusable templates so you stop retyping brand voice on every run — this is the spec-file idea, and it pays off with any tool.
Then watch the CrewTower thread. If the maker ships edit/cancel on queued prompts and smarter task-break detection, the pattern becomes genuinely production-grade — and I’d bet the first non-coding vertical to copy it will be video repurposing. The lesson for the rest of us isn’t “buy a notch app.” It’s that supervision is the next UX frontier for AI-assisted content, and the operators who make waiting states visible will out-ship everyone still hunting for the right terminal tab.
