The screen is part of the prompt — and creators are about to feel that shift
If you run social accounts for a living, you already know the dirty secret of AI tooling: the bottleneck was never the model. It’s the briefing. Every time I’ve tried to get a usable caption, hook, or edit note out of an AI assistant, I’ve burned more time describing what I’m looking at than I would have spent just writing the thing. Cakie, a new macOS beta from the maker of Shadow, is a direct bet on that bottleneck — it lets you record your screen, talk over it, point at things, and hand your agent a transcript plus the frames that actually matter. That’s a small-sounding feature that maps onto a very large problem for anyone producing video, carousels, and repurposed content at volume.
What Cakie actually is, and the problem it’s solving
Let me state the product plainly, then get to why it matters. Cakie is a macOS app, currently free in beta, built by Jay Song — the same maker behind Shadow. The pitch, in his own framing: he records his screen “like a Loom for your agent,” talks as if a teammate is sitting next to him, points or draws when precision matters, and then Cakie “picks the moments that matter.” A minute of video is thousands of frames, so instead of dumping the whole recording on an AI agent, Cakie uses your spoken words to select the few frames you actually meant. You review, delete screenshots you don’t want, and paste a single prompt into Claude Code, Codex, Cursor, or whatever agent you use.
The origin story is worth noting because it explains the design instincts. Song says he recently became a dad to twin girls, which ended his 16-hour days, and he now leans on AI agents for most of his product work. Voice typing sped up idea capture, but something was still missing: “A lot of what we communicate at work isn’t text. It’s the screen in front of us.” His canonical example is the kind of instruction every operator has typed and hated: “Make this Get started button bigger. no, not the one in the top bar, the second one, in the Team card.”
That’s a developer’s example, but the same failure mode shows up constantly in social work. Try describing a thumbnail composition to an AI image tool in pure text. Try explaining to an assistant which three seconds of a 40-minute podcast should become the hook. Try telling a junior editor which caption style on slide four is wrong without pointing at slide four. The screen is the context, and text boxes throw it away.
Why this is a creator problem disguised as a dev-tool launch
Here’s my take: Cakie is launching into the vibe-coding and agent-tooling conversation, but the people who should be paying closest attention are video-first creators and social teams. The reason is structural. Short-form video now dominates distribution on TikTok, Instagram Reels, and YouTube Shorts, and the operational reality of that dominance is volume. A single long-form asset — a podcast episode, a webinar, a product demo — is expected to spawn a dozen or more derivative clips. The repurposing workflow is where most teams drown, and it’s fundamentally a “point at the screen and explain” problem, not a “write a better prompt” problem.
If you can record yourself scrubbing a timeline, saying “this is the moment, cut here, this caption, this thumbnail,” and hand an agent that annotated recording as context, you’ve collapsed a 20-minute briefing into a 90-second one. That’s the promise. Whether Cakie delivers it is a separate question I’ll get to.
How it differs from the tools you’re already paying for
The obvious comparison is Loom, and Song invites it — “like a Loom for your agent.” But the comparison breaks down fast, and the break is the interesting part. Loom’s output is a human-facing video: a link a colleague clicks and watches. Cakie’s output is a machine-facing prompt: a transcript plus a curated set of screenshots that gets pasted into an agent. Those are different products that happen to share a recording UI.
The second comparison is to the AI clip-repurposing category — tools like Opus Clip, Descript, and CapCut’s various AI features. Those tools automate the editing: they find the clip, cut it, caption it, resize it. Cakie doesn’t edit anything. It automates the briefing. That’s a genuinely different layer of the stack, and I’d argue it’s the layer most teams have left manual. Opus Clip tells you which moment it thinks is good; Cakie lets you tell your agent which moment you know is good, with the screen as evidence.
The third comparison is to the scheduling and social-management incumbents — Buffer, Hootsuite, Later, Metricool. These tools own the distribution layer: queue, publish, measure. Cakie sits upstream of all of them, at the ideation-and-briefing layer. In a mature stack, you’d use Cakie to brief the clip, an editor or AI tool to produce it, and a scheduler to ship it. No overlap, which is a good sign for Cakie’s positioning — it’s additive rather than competitive.
Where the “picks the moments that matter” claim needs scrutiny
The load-bearing claim is that Cakie “uses what you said to choose the few you meant.” That’s a frame-selection problem, and it’s harder than it sounds. If you’re talking while you scroll, your speech and your screen are loosely coupled — you might say “this section is weak” while the timeline is already 30 seconds past the section. Song’s own example (“the second one, in the Team card”) is a static screen with a stable referent, which is the easy case. Video scrubbing is the hard case, and I have open questions about how well the transcript-to-frame alignment holds up when the screen is moving.
I’d also flag that the tool is macOS-only and free while in beta. Pricing after beta is not disclosed. For a solo creator on a Mac, that’s a low-friction trial. For a team standardized on Windows, or for an agency that needs SSO and admin controls, this is not yet a production tool. That’s not a criticism of a beta — it’s just the honest boundary of who should bother this week.
What social teams can steal from Cakie’s design, regardless of whether they adopt it
This is the section I’d actually want you to act on, because the most valuable thing about Cakie isn’t the app — it’s the workflow pattern it validates. Here’s what I’d borrow immediately.
Stop writing prompts. Start recording briefs. Even without Cakie, you can record a screen-share with voiceover using QuickTime or Loom, then feed the transcript plus a few manual screenshots into your AI tool of choice. It’s clunkier, but it tests the core hypothesis for free: does a spoken, screen-anchored brief produce better AI output than a typed prompt? In my experience testing similar approaches, the answer is usually yes for anything visual and no for anything purely verbal — captions and copy briefs don’t benefit much, but edit notes, thumbnail direction, and layout feedback do.
Build a “moment map” habit. The frame-selection idea is portable. Before you hand a long asset to a repurposing tool, spend five minutes marking the timestamps you care about and writing one line on why. That map is the single highest-leverage artifact in the repurposing pipeline, because it’s the difference between an AI tool guessing at virality and you asserting editorial judgment. Tools like Opus Clip will still pick their own moments, but you can override with yours — and your picks will usually win on brand fit even when they lose on raw hook strength.
Treat the screen as a first-class input in your AI stack. Most social teams have standardized on text prompts and maybe image uploads. Very few have a repeatable way to hand an agent motion plus annotation. That’s a gap worth closing, because the highest-value AI tasks in social — clip selection, thumbnail critique, layout QA, competitor teardown — are all inherently visual.
Why TikTok creators should care more than LinkedIn ones
A quick segmentation note, because “creators” is too broad a word. If your primary platform is LinkedIn, your content is mostly text and static images, and your briefing problem is mild — you can describe a post in a paragraph. Cakie is a nice-to-have. If your primary platform is TikTok, YouTube, or Instagram Reels, your content is motion, your briefing problem is severe, and your volume requirements are brutal. Cakie is closer to a must-try. The same logic applies to Pinterest for visual-heavy pin design and to Threads and X only marginally — text-first platforms just don’t stress the screen-as-context problem the same way.
Where my judgment says Cakie falls short
Let me be direct about the limitations, because a beta deserves honest scrutiny and because you’re deciding where to spend your attention.
First, it’s a briefing tool, not a production tool. Cakie gets context into your agent. It does not write your captions, cut your clips, schedule your posts, or measure performance. If you were hoping for an end-to-end social workflow, this isn’t it. The value is real but narrow.
Second, the agent dependency is a real constraint. Cakie’s output is “one prompt” for Claude Code, Codex, Cursor, or any agent. That means its usefulness is bounded by how good your agent is at acting on visual context. If your agent ignores the screenshots or misreads the transcript, Cakie’s careful frame selection is wasted. The tool is only as strong as the weakest link downstream, and that link is not under Cakie’s control.
Third, the beta caveats matter. macOS only. Free while in beta, with post-beta pricing not disclosed. No mention of team features, permissions, or shared libraries — which is fine for a solo maker but a blocker for agencies. And the maker explicitly asks for feedback on “where it gets things wrong,” which is a good sign of humility but also an admission that it will get things wrong.
Fourth — and this is my biggest open question — the frame-selection accuracy is unverified. Song claims Cakie “picks the moments that matter,” but there’s no published benchmark, no accuracy figure, and no independent testing I can point to. The claim is plausible given the design, but it’s a claim, not a demonstrated result. Treat it as a hypothesis to test, not a feature to rely on.
Where the math breaks
Here’s a concrete way the value proposition could fail. Say you spend 90 seconds recording a screen-share brief for a clip. If Cakie’s frame selection is accurate, you save maybe 10–15 minutes of back-and-forth with your agent or editor. If it’s inaccurate, you spend those 90 seconds plus the time to correct the output, and you’ve net-lost time versus just typing a prompt. The break-even depends entirely on accuracy, which is exactly the thing that’s unmeasured. That’s not a reason to skip it — it’s a reason to test it on a low-stakes task before you build it into your workflow.
What I’d watch / test next
If you want to put this to work this week, here’s what I’d actually do. First, pick one recurring briefing task you hate — clip selection from a long recording is the best candidate — and run it three ways: pure text prompt, screen recording with voiceover fed to your agent, and (if you’re on a Mac) Cakie itself. Time all three, and judge output quality blind. That’s a two-hour experiment that will tell you more than any launch post.
Second, watch whether Cakie ships Windows support, team features, and post-beta pricing — those three things determine whether it stays a solo-creator toy or becomes an agency tool. Third, keep an eye on whether the incumbents absorb the pattern. Loom adding agent-friendly transcript-plus-frame export, or Descript adding screen-recorded briefs, would be the strongest signal that “screen as prompt” is a real category rather than a clever one-off. My bet: it’s real, and the interesting question isn’t whether Cakie wins but whether the pattern becomes standard across the creator stack. Test it before you assume either way.
