Backed by Y CombinatorThree documents and two sheets that turn your best-performing writing into checkable patterns, not a page of adjectives.
Free download · No account needed
What you get
Pack of 5 files: 3 documents, 2 spreadsheets
Brand Voice Guide TemplateVoice Guide
Pulled from your own published writing, not described in adjectives.
Pattern ·
Before · from your own published work
After · the one identifiable fix
A pattern with no real pair behind it stays out of the guide.
A voice guideline built from adjectives asks two different writers to hit the same three words: confident, approachable, never salesy. Both drafts can defend both words, and neither is checkable against the other, because an adjective names a target without saying where the needle currently sits or which way to move it. One freelancer writes whatever confident means to them, the next one writes something else, also confident, and the brand reads like two different people within a week.
This pack replaces the adjectives with patterns pulled from your own best-performing writing, the pieces that were actually opened, read to the end, shared or converted rather than the ones somebody likes the sound of. Every pattern that survives carries a real before-and-after pair: an actual sentence you published, and the corrected version, per the rule Every Pattern Needs a Correction. A pattern with no real pair stays a hypothesis and does not go in on a guess.
The result installs beside the content brief that tells a writer what to cover and the editorial calendar that keeps drafts moving on schedule. Send River the 10 to 30 pieces you would call your strongest. Rank them by opens or clicks if that data exists, or say plainly that the selection is a judgment call. The Voice Guide and Word List come back derived from that writing, ready for a freelancer's first draft rather than their fifth revision.
Voice GuideFictional company: Thistlewood, project-management software. Corpus: 30 posts and 12 emails, narrowed to the 15 highest-engagement pieces.
Pattern 1 · second person, almost always
| Before | Users can now assign tasks directly from their inbox. |
| After | You can now assign tasks directly from your inbox. |
46 of 51 reader-facing sentences in the corpus said “you.” The 5 exceptions were one compliance email, not product copy.
Pattern 3 · one claim per sentence
| Before | Thistlewood helps you plan faster, communicate clearly, and ship on time, all from one dashboard. |
| After | Thistlewood plans your sprint. It flags what's off track before the standup does. |
The weakest piece in the corpus stacked three claims into one sentence; no top performer carried more than one past a period.
Word List34 terms total, sourced from the corpus's own weakest pieces and repeat review flags. Format: term, tier, reason, replacement.
| Term | Tier | Reason |
|---|---|---|
| unlock | don't use | Zero uses in the 15 strongest pieces; in 4 of the 5 weakest |
| seamless | don't use | Unfalsifiable; nothing in the corpus says what would make it “seamed” |
| simple / easy | caution | 8 of 11 uses described a task the writer had not actually tried |
| powerful | caution | Almost always stands in for a number nobody looked up |
21 of 34 terms are caution, 13 are don't-use. Three don't-use terms still stand inside direct customer quotes, which the checklist leaves untouched.
Review Log26 drafts run through the Review Checklist in the pack's first quarter of use.
| Check failed | Drafts | Share |
|---|---|---|
| 1. Word List | 18 | 69.2% |
| 2. One claim per sentence | 9 | 34.6% |
| 3. Address (you, not users) | 5 | 19.2% |
| 5. Section openers | 4 | 15.4% |
| 6. Terminology Register | 4 | 15.4% |
Word List violations led by a wide margin. “Powerful” and “robust” moved from absent to logged entries this quarter after the same reviewer caught the same pattern twice.
The patterns your best writing already shows, each backed by a real before-and-after sentence pair, never an adjective alone.
Terms your weakest work relied on, sorted don't-use or use-with-caution, with the reason specific to your own corpus.
Six checks against a new draft, in order, so nobody line-edits sentence six before check one has passed.
Product and feature names with the correct form beside the common error, checked last since it is the fastest pass.
One row per check that failed on a real draft, the record that eventually justifies a new Word List entry.
10 to 30 pieces, ranked by opens or clicks if you have that data, or the ones the team already agrees are the strongest voice.
River reads for what repeats: sentence length, address, contraction use, claim density, and pulls a real pair for each survivor.
The Review Checklist scores a current draft against the six checks, quoting the exact sentence that fails each one.
Every failure lands on the Review Log. A check that fails twice on different drafts becomes a candidate new guide entry.
Yes, the Word and CSV download needs no account or card. Edit with AI installs this pack in River and asks for your strongest writing before it drafts a single pattern. For getting that writing commissioned in the first place, see the content brief pack.
A tone deck states adjectives: confident, warm, direct. This pack states patterns instead, each one tied to a real sentence you published and, where it is a habit to avoid, a real sentence that broke it. A new writer can extend a documented pair to a sentence the guide never saw, which a personality description cannot support.
Send whichever pieces the team already agrees are the strongest examples of how you want to sound, and say that it is a judgment call rather than a ranked list. The patterns still get pulled the same way; only the input changes from measured to agreed-upon.
No. Every entry traces to this corpus specifically: a term found in the company's own weakest pieces, or corrected twice in real review. The two-tier structure follows Google's own developer style guide, which separates use with caution from don't use with a stated reason per entry rather than a flat ban list.
That is the more common use once the guide exists. Point a draft, AI-written or not, at Review a Draft and it runs the same six checks either way. SEO writing review still checks a draft against its brief; this pack checks whether it sounds right.
Only when a correction repeats. Promote a Pattern from the Review Log looks for the same check failing on two or more different drafts. It proposes the new Word List entry or Voice Guide pattern from those real rows, rather than updating on a single reviewer's one-off preference.
Written by
Co-founder & CEO, River
Chandler has spent the last six years working in AI, building tools for millions of users and Fortune 100 companies. In 2025, he co-founded River to build a new kind of operating system for AI agents.
Send your strongest pieces and River pulls the patterns, each one backed by a real before-and-after pair.
Edit with AI