It Waits for You

Your turn to talk

Growl reads each step, then listens for you. The meter shows it's ready, so you can answer from across the kitchen without looking for a button.

One switch to go hands-free

Turn listening on once, and Growl listens after every step for the rest of the cook. Every command is on the same sheet, so there's nothing to memorize.

Why Growl takes turns

The narration uses everyday words like "back" in 41 of 1,340 steps. Listening between steps means the only voice Growl answers to is yours.

Say it like you would to a friend

Six everyday words cover the whole cook: next, back, repeat, pause, resume, and timer. Say "stop" and Growl simply pauses, so you never lose your place.

Ready when you are

The meter rests while Growl reads the step and comes alive when it's your turn. You always know when to speak.

Per Step Timer Sync

Works with “Timer” speak on any step or manually as well.

Describe what belongs

Telling the model to leave shrimp heads out made them more prominent every time. So every prompt describes only what should be in the frame, and nothing that shouldn't.

One cause behind every miss

In one batch, all 13 problems I found by eye had the same root: the model filled any gap with the most typical version of the scene. Fixing that one cause fixed the batch.

Nothing generates without a yes

Every prompt passes an automated check, then my review, before anything is made. Afterward, a review tool flags the steps most likely to need a closer look, and it caught every known issue in the first batch.

Some things are better shown in two frames

After a dozen attempts at a clean knife cut, I stopped generating blades in motion entirely. Cuts now show the before and after as two stills, which is clearer to follow and costs a third as much.

Four tiers, one score

Exact, word-prefix, substring, typo, with name outranking chef, tags, and ingredients. The weak tiers earn their place: dropping them loses "virgin" finding extra-virgin olive oil.

Suggestions never guess

Typo tolerance runs in results but never in suggestions, so the bold typed prefix and grey remainder always describe a real match. Three other suggestion budgets were tried first: one shared cap hid every ingredient, separate quotas pushed the list below the fold.

When nothing matches, trim the query

"Adfaes" becomes "Ad" and finds Chicken Adobo. Edit distance over the whole catalog was rejected because it resolves to whichever names are shortest, and the general matcher returned Avocado and Pad Thai.

Short words never get typo tolerance

"sand" matched dozens of recipes through the word "and," so words under four letters match literally or not at all. Autocorrect is off in all four fields for the same reason: it turned "Borsc" into "Borax."

Watch & Cook, not a video toggle

122 of 173 recipes are audio. A two-sided filter would publish the production economics as a quality tier and mark most of the catalog as the lesser half. So video gets a door, not a switch.

Step by Step Audio

Most steps you just need to hear. Audio narrates the step over its text, with a small clip looping beside it, so you can follow along without watching the screen.

Step by Step Video

Some steps you need to see. Pleating, rolling, and forming get full-screen video, so the technique plays out in front of you while the narration talks you through it.

A quarter of the cost

Measured on the first audio recipe, Hummus at 11 steps, the audio version cost 32 credits against about 140 as video. That difference is what makes a catalog of a thousand recipes possible.

The score informs. It doesn’t decide

The score predicts which steps are hardest to produce. Those are the steps with real technique in them, which is exactly when a cook needs to watch, so they stay video even when they cost more.

Narration sets the pace

Audio is generated first, because its length decides how long a step runs. Video loops silently underneath, and the step only moves on when the narration is done.