The advertised feature is never the real constraint
3 min liveSession 2 gave you prices. Prices are the easy part, because vendors publish them loudly and in one place. What they do not publish loudly is the sentence that decides whether the feature works on your footage at all.
Across every vendor in this category, the real constraint sits about three clicks deeper than the page that sold you the tool. Not hidden, not illegal, not even unusual. Just filed one level down, in a help article that the marketing page does not link to, written in a flat voice that assumes you already committed.
Here is the version of that sentence you will meet tonight, quoted exactly: Studio Sound "only works on recorded audio, not AI-generated speech". Remove Filler Words "currently detects filler words in English transcripts only". Eye Contact is "single person only: the effect works when one clear face is in frame". Regenerate "currently supports English audio and video only". None of those four sentences appears anywhere near where you make the buying decision.
LiveWhy this is a lesson and not a list of gripes2 min▶
A list of limits goes stale in a quarter. Every one of the quotes above will be edited, renamed or superseded, probably before this course is a year old, and session 1 showed you exactly how fast that happens.
What does not go stale is the shape of the limits. Read enough of these pages and you stop seeing twelve unrelated footnotes and start seeing five recurring gates: a language gate, a count gate, a duration gate, a scope gate and an expiry gate. Every constraint in this session is one of those five wearing a different product name.
That is the portable part. A reader who knows the five gates can open the docs of a tool this course has never mentioned, find the real limit in about ninety seconds, and decide before paying. That is the whole point of tonight, and everything else on this page is evidence for it.
Five gates, and the question each one answers
7 min liveLearn this diagram and you can stop memorising vendor footnotes. Each row is a question, and the docs of any AI media tool will answer at least three of them if you go looking.
Notice what the five gates have in common. None of them is about quality. None says the output will look bad. Each one says the feature will simply not engage, or will engage on something other than what you pointed it at. That is a much worse failure than a mediocre result, because a mediocre result tells you it happened.
It is also why the gates are worth memorising rather than looking up. You will meet them at the moment of highest pressure, which is the moment you have already recorded the thing and already promised the delivery.
Self-studyWhy vendors file it this way, without anybody being dishonest3 min▶
It is tempting to read this as deception. Mostly it is not, and the more accurate story is more useful to you.
A feature page has one job: explain what the thing does to somebody who has never seen it. Adding "single person only, and it fails on variable frame rate footage" to that page would be honest and would also make it unreadable. So the limit goes to the help article, which is written later, by a different team, for somebody who already has the problem.
The result is structural rather than malicious, and it is completely predictable. Which means you can route around it. The habit is one line long: before you buy or before you commit a deadline, open the help documentation for that exact feature and read to the bottom. The bottom is where the gate lives, every time, because that is where the writer put the caveats they could not fit above.
What the page says, and what the docs say
7 min liveThree features, three pairs of columns. The left column is the promise you formed in your head. The right column is quoted from the vendor's own help pages, and nothing in it is disputed by anybody.
Take the Studio Sound row seriously, because it contains the worst outcome on this page. The documentation says that with extremely loud background noise the model "may not be able to separate the speech from the noise, resulting in a silent waveform". Read that again. The failure mode of an audio cleanup tool is that it hands you back silence.
And because the same feature is scope-gated to the file, not the clip, you cannot switch it on for the one noisy interview and leave the rest alone. You are making a decision about the whole file. Those two facts are on the same help page and neither is on the feature page.
Self-studyThe filler word count that people report as a bug4 min▶
This one is worth knowing because you will be asked about it, and the intuitive answer is wrong.
Remove Filler Words shows you a count of what it detected. It then removes fewer than that, and people file bug reports. It is not a bug. The option called "avoid harsh cuts" deliberately skips filler words that cannot be removed without clipping into nearby words or leaving an awkward cut behind.
So the removed count will never equal the detected count, and it should not. A tool that removed all of them would be a tool that clipped the first syllable off the word after every skipped "um". The gap between the two numbers is the feature protecting your audio.
The other constraint on the same feature is the language gate: it "currently detects filler words in English transcripts only". If you work in any other language, this feature does not exist for you, and it will not tell you that. It will just find nothing.
Self-studyEye Contact, and the remedy the docs recommend3 min▶
Eye Contact is a count-gated feature and states it plainly: "single person only: the effect works when one clear face is in frame". Two-hander interviews, panels, anything with a second person drifting into shot, all out of scope. That is a hard gate, not a quality warning.
The second gate is technical and catches far more people. It fails on sideways footage, and it fails on variable frame rate video. Variable frame rate is what a lot of screen recorders and phone capture apps produce by default, which means a large share of the footage people actually want to fix is exactly the footage the feature refuses.
The tell that this is a real limit rather than a temporary one: the documentation's own suggested remedy is to re-encode your file in HandBrake first. A vendor sending you to a third-party transcoder before its own feature will run is a vendor telling you, politely, that this is not getting fixed soon. Read that as a signal whenever you see it.
Two seconds is the whole feature
5 min liveDuration gates are the easiest to check and the most often skipped, because a number that small does not feel like it can be the real answer. Here is Adobe Generative Extend drawn to scale.
The asymmetry is the interesting part. Audio gets five times the allowance video does, and that is not arbitrary. Extending room tone convincingly is a far smaller problem than inventing two seconds of a person continuing to move. The gate is telling you where the technology actually is, if you read it as evidence rather than as a limitation.
There is a second gate stacked on top: it is cloud only, and it requires a minimum of two seconds of source video or three seconds of source audio before it will engage at all. So the feature has a floor and a ceiling, and the usable window between them is narrow enough to plan around. Which is good news, as long as you know the numbers before the shoot rather than after.
Self-studyExpiry gates, the one that is invisible until it fires3 min▶
The other four gates announce themselves at the moment you try to use the feature. Expiry gates do not. They announce themselves later, usually when you go back for something.
Opus Clip's free tier states it in one line: "After 3 days, the clips will no longer be exportable." Not deleted. Not hidden. Not exportable. You can still see the work you did, which is arguably worse, because the thing that is gone is the only part that mattered.
Treat any free tier as expiry-gated until proven otherwise, and go and find the sentence. The question to ask is not "is it free" but "for how long can I get my work back out". Those are completely different questions and only one of them is answered on the pricing page in large type.
Two renames that broke every tutorial about them
5 min liveA gate you can read is a solvable problem. A gate you cannot find because the feature changed its name is a different kind of problem, and it is the reason most third-party material about these tools is now misleading rather than merely dated.
"Overdub" no longer exists. The word appears zero times across Descript's complete 398-entry documentation index. The feature is now AI Speech, AI Speaker, or a custom voice clone depending on where you look, and the Terms of Service section governing it is headed "Use of AI Speakers". Meanwhile the legacy Overdub blog posts are still live on the marketing site, which is precisely why third-party courses are stale without knowing it.
The Adobe one is quieter and just as awkward. The product is now documented as "Adobe Premiere", while the URL paths still contain the string "premiere-pro". So the name you search for and the name the documentation uses have separated, and you can land on a current page while believing you are looking at an old one.
Self-studyThe vendor candour worth copying3 min▶
It would be unfair to leave this session sounding like every vendor buries everything. One of them writes a warning so blunt it is worth quoting as a model of what good documentation looks like.
Descript, describing the output of ElevenLabs v3: "You may notice inconsistencies in how the output from this model sounds across paragraphs. This can include shifts in tone, volume, accent, or even speaker identity."
Sit with the last three words. The vendor is telling you, in its own help pages, that the synthetic voice may stop sounding like the same person partway through. That is not a small caveat, and putting it in writing costs them something.
Use it as your calibration. When a vendor writes at that level of candour, the rest of its documentation is probably trustworthy. When a feature page has no limitations section at all, that absence is itself information, and session 5 shows you a system card that has exactly that shape.
Find the gate in a tool you already pay for
★ 18 min · everyone readsNo simulator tonight. The skill is a reading habit, and the only way to acquire it is to do it once, slowly, on a tool you actually use. Pick that tool now, before step one.
Feature page. This is where you formed your belief about what the tool does. Write that belief down in one sentence, now, before you read anything else. You will want the exact wording later.
Pricing page. Tiers, credits, media minutes. Useful, and it is session 2's territory. What is almost never here is whether the feature works on your footage.
Documentation page for that one feature. Not the docs home. The article about the single feature you care about. Scroll to the bottom third and read every sentence that starts with a qualifier.
The gate. Language, count, duration, scope or expiry. There is usually more than one. Write them next to your sentence from step one and look at the distance between them.
Name your tool and your one sentence. "It cleans up bad audio in one click." "It makes me look at the camera." Whatever you actually believe. The point of writing it down is that you cannot quietly revise it once you have read the docs.
Open the help documentation for that exact feature, not the product. Search the vendor's own help domain. If your search returns marketing pages and blog posts, add the word that appears in their navigation, usually "help" or "docs". This is where the Overdub problem bites: search the current docs index, never the blog.
Scan for the five gate words. Language, English, only, single, maximum, minimum, seconds, per file, per clip, days, expires. Any one of them is a gate. Highlight every sentence containing one.
Classify each one. Language-gated, count-gated, duration-gated, scope-gated, expiry-gated. Most features have two or three at once, and the combination is what actually decides whether the feature is usable for you. Eye Contact is count-gated and format-gated together, and either one alone would be survivable.
Now check your sentence from step one. Would you still write it? For most people the honest answer is that they would add three words, and those three words are the ones that would have changed a purchase decision or a deadline promise.
Test the worst gate on real footage before you need it. Take thirty seconds of your actual material and run it through. This is the whole exercise in one line: you want the feature to fail on a Tuesday afternoon with nothing at stake, not on the evening of the deadline.
Why this is worth eighteen minutes rather than a bullet list. Everyone agrees in principle that you should read the documentation. Almost nobody does it before buying, because the feature page is persuasive and the docs page is boring. Doing it once, on a tool you already own, converts a piece of advice into a reflex. The reflex is the thing you are buying with tonight.
Before session 4
30 min- Write the gate card for the one tool you would miss most. Five lines, one per gate type, each answered or marked "not stated". The "not stated" lines are the valuable ones, because they are the questions you will have to answer by testing.
- Deliberately trigger one gate. Run a filler-word pass on non-English audio, or an Eye Contact pass on a two-person clip. Watch exactly how it fails. Most gates fail silently, and knowing what silence looks like in your tool is worth more than the reading.
- Search the vendor docs for a feature name from any tutorial you have followed. If it returns nothing, do not assume the feature was removed. Look for what it was renamed to. Overdub is the reference case: gone from the docs, alive on the blog.
- Find one vendor sentence that is unusually candid. The ElevenLabs v3 note about speaker identity is the standard to measure against. A vendor willing to write that is a vendor whose other numbers you can probably trust.
- Optional, and it will change how you buy: take any AI feature you have not paid for yet and try to find its constraint in under ninety seconds. If you cannot find one at all, that is not reassurance. It means the limitations section does not exist, and session 5 shows you what that usually indicates.
Three questions
Clear all three and session 3 gets stamped in your passport.
1 · You run Studio Sound over an interview recorded in a loud cafe, and the result comes back as a silent waveform. What happened?
This is on Descript's own help page for the feature. Two things make it worse than an ordinary bad result. It destroys your audio rather than degrading it, and because Studio Sound is scope-gated to the file rather than the clip, you were never able to apply it to just the one interview in the first place.
2 · Remove Filler Words reports 41 detected and 33 removed. A colleague wants to file a bug. What do you tell them?
The gap between the two numbers is the feature protecting your audio. A tool that removed all 41 would be a tool that clipped the first syllable off the word after every skipped filler. Worth remembering separately: the same feature is language-gated to English transcripts only, and in any other language it does not report a problem, it just finds nothing.
3 · A tool this course has never mentioned launches an AI editing feature next month. What is the first question you ask?
Price is session 2 and it is genuinely easy to find. The model lineage is interesting and rarely decisive. The gate is the thing that determines whether the feature engages on your footage at all, and it is the one answer the marketing page will not give you. Learning the five types is what makes this portable to tools that do not exist yet.
What this session teaches, and where it came from
Every fact on this page, and its verification tier, is recorded in the course's source map., including a quarantine list of what is deliberately not stated.
Session 3 cheat sheet · pin this
Tonight every constraint was a vendor's choice, and a vendor can change its mind. Session 4 is about the one requirement nobody in this stack gets to set, where "we have transcripts" turns out to be a Level A claim being sold as Level AA, and where the automatic captions you were about to ship score an average of 7 percent error and a worst individual file of 53.8.