I thought I was working on content. I was actually changing input data for a feature.
RYZA has premium intention packs: short sentences a user reads aloud to dismiss an alarm or move through Wake Sequence, the short follow-up challenge that runs after the alarm stops ringing. I had 480 premium sentences, covering focus, courage, consistency, recovery, goals, and a few other moods people might want in the morning.
On paper, that sounded like plenty. Then I used them as voice missions.
Some were too long. Some sounded like things no one would naturally say out loud at 6:30 a.m. Some were technically valid sentences but broke badly when turned into Wake Sequence cards. A few were short enough to look good in the app, but too vague for speech recognition to handle consistently.
The collection read fine on a screen. It just was not written to be spoken into a phone while an alarm keeps going.
That difference changed the job.
A sentence can be good and still fail the mission
A good sentence and a good voice mission are not the same test. The earlier set had plenty of lines that passed the first and failed the second, one of them being:
“I make the next step smaller, then I take it.”
Nothing is wrong with that sentence. It is calm, grammatical, and easy to read. But a voice mission has more constraints than a quote card. The sentence needs to be:
- easy to say before the user is fully awake;
- distinct enough for speech recognition;
- short enough to repeat without frustration;
- natural in each supported language;
- usable as a sequence if the user chooses Wake Sequence.
That last part caused more trouble than I expected. Wake Sequence can split a sentence into smaller cards. A phrase that looks elegant as one line can turn into awkward fragments. In English, spaces make splitting seem easy. In Korean and Japanese, that assumption gets dangerous fast.
I had been judging the text with my eyes. The app was asking people to use their mouths, ears, and sleepy attention.
I started by making sentences shorter. That was not enough.
My first pass was mostly editing for length. I removed filler, cut long clauses, and changed some abstract encouragement into direct statements. That helped, but I still found sentences that were short and annoying.
A short phrase can be too generic, for example:
“I move forward.”
It is easy to say. It is also close to a lot of other sentences. Speech recognition can hear several variations of it, and the user does not get much sense of what they are actually saying.
Other short phrases were too compressed:
“Plan into practice, now.”
That may look sharp on a card. Spoken aloud, it can sound unfinished. Some people will like it. Others will hesitate over where the sentence is supposed to land.
I needed a different standard. The question became: would I say this sentence out loud once, clearly, without feeling like I am reading copy? That is not a perfect test, but it was more useful than counting characters.
I gave the catalog a mission contract
A mission contract is just a set of rules the content has to satisfy before the app accepts it. For the premium catalog, I started treating each sentence as something with a job, not just a message. It had to fit its category, stay within a reasonable length, remain speakable, and have a safe segmentation shape for its language.
It never turned into a single style guide. What I kept coming back to was closer to an awkward checklist:
Does it sound like a person could say it?
Does it have enough concrete words to recognize?
Would I be annoyed to repeat it three times?
Can its pieces still make sense in a sequence?
Does the translation still mean the same thing?That led to sentences with more texture. Instead of “I make progress visible with one completed action,” I wrote things closer to “Finished work is the only kind anyone can see.” Instead of “I give my best attention to one important thing,” I used “One thing gets everything I have.”
I would not call these universally better sentences. They are more specific, though, and that gives the voice mission something firmer to hear.
Some of the rewritten lines are deliberately plain: “I pick one and go.” “The day is not over yet.” “Fear rides along, I drive.” None of them are trying to sound like a meditation app. The goal is just that they get said.
The content rewrite exposed an uncomfortable gap
After rewriting the 480 premium sentences, I found another problem: some of the free sentences were still hard to pass as voice missions.
That was frustrating. I had just finished a much larger rewrite, and now the product was telling me that the small set needed attention too.
I added 58 short sentences across the catalog to give the mission more reliable options. I did not add them for a release note line. I added them so users would stop landing repeatedly on phrases that were technically supported but hard to actually complete.
A premium catalog cannot be carefully designed while the free starting experience is left behind. Users meet the mission before they understand the catalog structure.
Japanese made the hidden problem obvious
The biggest regression appeared in Japanese. Wake Sequence was splitting text into cards, and a change in the segmentation behavior caused Japanese cards to break in ways that looked unnatural or made the mission harder than intended.
This was the part where I had to admit my earlier assumption was shallow. I had treated segmentation as a utility: take text, break it up, show cards. But card boundaries are product behavior.
A bad boundary can change the meaning, make a phrase difficult to read, or turn a fair wake-up task into a guessing game. Japanese made that visible quickly because the usual “split on spaces” instinct was useless.
I fixed the regression and added mission-contract tests around the catalog. The tests do not prove that every sentence is beautiful. They check that the content still fits the structural rules the feature depends on.
That is a different kind of quality control from proofreading.
I corrected more than wording
During the full audit, I also reviewed the intention copy across five languages and corrected hundreds of places where the wording was awkward, too literal, inconsistent, or simply not the tone I wanted.
There were 669 corrections in that pass.
I do not think that number is impressive by itself; mostly it means the earlier catalog had a lot of small debts hiding in it.
Some changes were grammatical. Some made a sentence easier to say. Some removed the kind of generic encouragement that sounds fine until it appears at 6 a.m. on someone’s lock screen.
I kept finding the same pattern: a line could be correct and still not belong in the product.
What I got wrong
I had treated the premium sentence library as a finished collection because it was large. It was not finished. It had not been tested in its real environment.
A content catalog becomes software the moment a feature depends on its structure. If it is read aloud, it needs to work aloud. If it becomes cards, its divisions need to work. If it exists in five languages, “translated” is not the same as “ready.”
The rewrite did not make all 480 sentences perfect. I still expect to replace individual lines as people use them. There are phrases I like on paper but would probably remove if they repeatedly cause recognition trouble.
But the catalog stopped being a shelf of motivational quotes. Now it is part of the alarm interaction.
That made the editing slower. It also made the product more honest.