The Real Product Lesson in Foley Has Nothing to Do With To-Do Lists
Every social media operator I know runs their day inside a tool that was designed for someone else. We schedule in one app, write hooks in another, track links in a third, and keep the actual editorial calendar in a fourth — usually a Notion doc that’s three weeks out of date. So when a maker launches a voice-first task capture tool and the comment thread turns into a debate about when an AI should ask a clarifying question instead of guessing, my ears prick up. That’s not a productivity-app conversation. That’s the exact failure mode that kills AI-assisted content workflows.
Foley is a voice-to-task app from José Airosa, a software engineer who says he has over 25 years of experience. You talk, it files. It parses relative time (“next Tuesday morning”), it asks when it’s unsure, and it moves tasks when your plans change. The desktop app is free with no account if you bring your own OpenRouter key; Foley Pro is $8/month on the web, with the first month at $4 applied at checkout. That’s the whole pitch. On its face, it’s a to-do list. In practice, the launch thread is a masterclass in the single hardest problem in AI content tooling right now: deciding when a model should act and when it should stop and ask.
That’s why I’m writing about it here rather than in a productivity newsletter. If you’ve ever had an AI tool schedule a post for the wrong day, hallucinate a caption in the wrong brand voice, or auto-reply to a comment with something that made you wince, you already understand the stakes. The Foley thread is a rare public look at how a solo builder is drawing that line — and the tradeoffs he’s hitting are the same ones you’ll hit the moment you wire an LLM into your publishing pipeline.
What Foley Actually Solves (and Why the To-Do Framing Undersells It)
The origin story is mundane and therefore credible. Airosa says he kept losing tasks he said out loud — mid-meeting, in the car, walking the dog. That’s the same friction every creator knows: you have a hook idea in the shower, a reply you owe a brand partner while you’re driving, a video concept while you’re on a walk. By the time you open your phone, unlock it, find the right app, and type, the thought is gone.
Foley’s bet is that the capture step should be voice-first and that the filing step should be handled by a model that understands human time. “Next Tuesday morning” is the canonical example, and it’s a genuinely hard parsing problem because it’s ambiguous in at least three ways: which Tuesday, what counts as morning, and whether “next” means the upcoming one or the one after. The maker’s answer, per the thread, is that Foley doesn’t trust the model with the calendar at all. When it resolves a weekday, the app checks the date in code against your device’s clock before saving. That’s the right instinct, and I’ll come back to why in the limitations section.
The second bet is the clarifying question. Airosa’s line — “A good question beats a bad guess” — is the thesis of the whole product. It only asks when it’s genuinely unsure: two people named Sam, a project it hasn’t seen, or a reminder with just “morning” and no time. Otherwise it files silently, and you see what it understood before the task lands so you can correct it there.
The self-correction problem is the interesting one
A commenter from Refocus, who says they’re building voice AI for older adults, raised the case that breaks most voice capture: mid-sentence self-corrections. “Thursday, no wait, Friday.” Does Foley treat the last thing said as final intent, or does it ask you to confirm?
Airosa’s answer is that Foley reads the whole sentence before filing, and the last value wins — so that example becomes one task on Friday, no extra confirm step. He also volunteered the hardest real miss they hit: “Add a task to call the dentist tomorrow. Update it: Friday.” The word “Update” pulled the model toward editing an older task instead of modifying the one it had just created. They fixed it in the task structure, not the prompt, because — his words — changing the prompt alone didn’t stop it.
That detail matters more than the product itself. If you’ve ever tried to get an LLM to reliably output structured data for a content calendar, you know the prompt-only fix is a trap. You tweak the system prompt, it works for a week, then a slightly different phrasing breaks it. The durable fix is almost always schema-level: constrain the output shape, validate it in code, and only then let the model’s judgment apply inside that box.
The “When Should AI Ask?” Debate Is the Whole Ballgame for Content Teams
The most substantive exchange in the thread is between Airosa and Gal Dayan of Dial. Dayan’s opening line is worth quoting because it’s the cleanest statement of the problem I’ve seen from a builder: “asking feels like friction and guessing feels like magic, until the guess is wrong and now you’ve got a task filed under the wrong project with a made-up date.”
That’s the exact dynamic in AI content tooling. The reason Buffer, Hootsuite, Later, and Metricool have all bolted AI features onto their schedulers over the last two years is that “generate a caption” feels like magic in a demo. It stops feeling like magic the first time the model writes a caption in the wrong tone for a client account, or schedules a product launch post for a Sunday at 3am because it didn’t understand your timezone. The friction of asking — “which account?”, “what tone?”, “when exactly?” — is annoying right up until it saves you from a mistake you’d have to explain to a client.
Dayan then pushes on the mechanism: is “not sure” a confidence-score threshold, or ambiguity between two specific candidates? Airosa’s answer is that he lets a model judge ambiguity rather than hand-tuning a score, and he name-drops that this came together around the time Jev launched — though the source doesn’t give more detail on what Jev is or how it’s integrated, so I won’t speculate.
Dayan’s follow-up is the sharpest question in the entire thread, and it’s the one every operator should be asking about their own AI stack: a model’s sense of “clear enough” can flip between two phrasings of basically the same request. Have you noticed Foley asking on one wording and guessing on a reworded version of the same thing?
Airosa concedes the point directly. He says consistency is hard to guarantee because models are inherently non-deterministic, that there’s no guarantee of the same outcome every time, and that he supports the model with “a good amount of good old-fashioned code” to attempt consistency. Dayan’s final reply — which appears to be the last message in the scrape — asks whether that code layer catches repeats of the same ambiguous request (so the second time someone phrases a task a certain way, it remembers how it got resolved) or whether it’s a hard keyword trigger that forces a clarifying question regardless of what the model decides. As of the scrape, that question is unanswered.
Why this unanswered question should bother you
If you run any AI-assisted workflow at scale — caption generation, comment triage, repurposing long-form video into shorts, drafting LinkedIn posts from a podcast transcript — you are living inside this exact ambiguity. The difference between “the model remembers how we resolved this last time” and “the model re-rolls the dice every time” is the difference between a tool you trust with your brand voice and a tool you babysit.
The first option is essentially a learned preference layer. The second is a guardrail. Both are useful, but they solve different problems, and most AI content tools on the market today don’t clearly tell you which one they’ve built. That’s a question worth asking vendors directly.
Why TikTok creators should care more than LinkedIn ones
Here’s my take, and I’ll flag it as opinion: the tolerance for AI guessing wrong is wildly different by platform, and by extension by the teams that run those platforms.
On TikTok, speed is the whole game. A hook idea captured in ten seconds and posted the same day is worth more than a perfectly polished idea that ships next week. If your AI capture tool asks one clarifying question and gets the task filed correctly, that’s a net win — the friction is trivial compared to the cost of losing the idea. On LinkedIn, the calculus inverts. A single mis-scheduled post or a caption in the wrong voice can cost you a client relationship, because the audience is professional and the stakes are reputational. If you’re running both, you want the same underlying tool to behave differently depending on which account the task is bound to — which is a feature almost nobody has built yet.
What Creators Should Steal From This Thread (Regardless of Whether You Use Foley)
You don’t need Foley to benefit from the design thinking in its launch thread. Here are the transferable moves, in rough order of how much they’ll save you.
1. Separate the model’s job from the code’s job. Airosa’s calendar check is the template. The model interprets intent; the code validates anything that has to be objectively correct — dates, times, timezones, character counts, hashtag limits, UTM parameters. In your own stack, that means: let the LLM draft the caption, but validate the scheduled time, the platform’s character limit, and the link’s UTM string in code before anything publishes. Zapier and Make both let you insert validation steps between an AI action and a publishing action, and if you’re on n8n you can do it natively.
2. Make the correction step visible, not hidden. Foley shows you what it understood before the task lands, so you can fix it in place. That’s the difference between an AI tool that earns trust and one that erodes it. When you build or buy AI caption tools, favor ones that show you a draft you can edit over ones that publish directly. This is also why CapCut’s auto-captioning works for so many creators despite being imperfect — you see the transcript and can fix it before export, rather than discovering the error after the video is live.
3. Fix structural problems structurally. The “Update it: Friday” bug is the lesson. When your AI pipeline fails on a specific phrasing, resist the urge to patch the prompt. Ask instead: is there a schema change, a validation rule, or a routing rule that makes this class of failure impossible? Prompt patches are band-aids. Schema changes are surgery.
4. Treat relative time as a first-class problem. If you schedule content across Instagram, TikTok, YouTube, X, LinkedIn, Facebook, Threads, and Pinterest, you already know that “post this tomorrow morning” means eight different things depending on the platform’s peak hours and your audience’s timezone. Any tool that accepts natural-language scheduling needs to resolve that ambiguity explicitly, not silently.
Where the math breaks
There’s a cost dimension here that the thread doesn’t address, and I’d bet it matters for anyone considering the free desktop tier. Foley’s free desktop app requires your own OpenRouter key, which means you’re paying per token for every voice capture that gets parsed. For a personal to-do list, that’s pennies. For a content team capturing hundreds of ideas a week across multiple accounts, the token cost of every “is this ambiguous?” check adds up — and the clarifying-question design, which is the product’s best feature, is also the thing that multiplies your API calls. Airosa doesn’t disclose token costs or usage limits in the thread, so I can’t tell you what this actually runs. But it’s the first question I’d ask before rolling it out to a team.
Where I Think Foley Falls Short — and Who Should Skip It
I want to be balanced here, because the thread is genuinely good and the maker is unusually candid about failure modes. But there are real gaps.
It’s a single-player tool in a multiplayer job. Social media is a team sport. Even a solo creator usually has a VA, an editor, or a brand partner in the loop. The thread says nothing about shared lists, assignments, comments, or permissions. If you’re running content for a client, you need a tool where the client can see status without you forwarding screenshots. Foley, as described, is a personal capture tool. That’s a real limitation, not a knock — but it means it’s not replacing your project management layer.
The consistency question is genuinely open. Airosa’s own answer concedes that the model’s judgment can flip between rewordings, and that the code layer is the mitigation. Dayan’s final question about whether that layer learns from prior resolutions is unanswered in the scrape. Until it’s answered, I’d treat Foley as a tool that’s usually right and occasionally surprising — which is fine for personal tasks and risky for anything client-facing.
No integrations mentioned. The thread mentions an Apple Watch capture flow, which is a nice touch for the “phone is too much friction” case. But there’s no mention of calendar sync, Slack, Notion, or any of the tools creators actually live in. For a to-do app in 2025, that’s a notable absence, and it’s the thing that would keep me from adopting it as my primary capture tool.
Who it’s not for: If you need team collaboration, client-facing reporting, or tight integration with your existing content stack, Foley isn’t it yet. If you’re a solo operator who loses ideas between the car and the desk, the free desktop tier is a low-risk experiment — assuming you’re comfortable managing an OpenRouter key.
What I’d Watch / Test Next
Three concrete things I’d do this week, whether or not you try Foley.
First, audit your own AI tools for the ask-vs-guess line. Pick one AI feature you use weekly — caption generation, comment replies, video repurposing — and deliberately feed it the same request phrased three different ways. If it behaves differently each time, you’ve found the inconsistency problem in your own stack, and it’s worth adding a validation step or a template.
Second, if you try Foley, test it against the self-correction case. Say “post this Thursday, no wait, Friday” and see whether it files one task or two. Then say “Add a task to film a reel tomorrow. Update it: Friday” and see whether it edits the new task or an old one. That’s the exact bug the maker says they fixed, and it’s the fastest way to judge whether the fix holds.
Third, watch whether Airosa answers Dayan’s final question. If the code layer learns from prior resolutions, that’s a meaningfully better product than one that re-rolls every time — and it’s the kind of feature that would make Foley worth recommending to creators who capture ideas on the move. If it’s just a keyword trigger, it’s a decent guardrail and nothing more. Either way, the answer tells you something about how seriously the maker is treating the hardest problem in AI-assisted workflows: knowing when to stop and ask.
