Voice Library: bootstrap VibeFlow with your past work

The four-generation valley

There's a quiet failure mode that almost every AI marketing tool ships with, and almost no AI marketing tool talks about: the cold-start valley.

You sign up. Your first three or four generations are mediocre — generic SaaS phrasing, off-voice, slightly too long, all the AI tells. By generation seven or eight, the tool starts catching on to your taste. By fifteen, it sounds genuinely like you.

That's a great curve. The problem is half the new users churn before they get there.

Phase 2 of our Learning Engine, which I shipped two days ago, addresses the long-run version of this — every campaign you generate makes the next one tighter. But it didn't fix the first three or four. New users still walked into a four-generation valley before they felt the personalization.

We just closed the valley.

What we shipped today

VibeFlow's new Voice Library lets you upload your past work — blog posts, emails, social threads, anything you've written that already sounds like you — and we use it as a voice anchor on every generation. Paste five things. Two minutes of work. Your first campaign now reads like generation fifteen.

The mechanics are simple:

• Paste text directly, paste a URL (we fetch and parse the body), or skip and add later. .docx and .pdf are coming next week.
• The library lives privately on your account. RLS-locked, never shared, never used to train cross-user models.
• At generation time, your two most-recent uploads get folded into the prompt preamble alongside your prior VibeFlow campaigns. Same flow, more signal, no extra search credits.
• If you skipped during signup, an inline paste box on the dashboard catches you. Dismiss it once and it stays out of your way; come back when you're ready.

You can manage the library any time from Settings. Add, browse, delete. The model of "you" updates on the next generation.

Why this matters more than it sounds

The feature itself is small. The shift it represents is bigger.

Most AI marketing tools build their personalization on top of future user activity — they learn from what you generate inside the tool. That's a fine model if your users stay long enough for it to matter. But it leaves the most expensive question — does VibeFlow sound like me on day one? — answered with "kind of, eventually."

The Voice Library inverts the assumption. We learn from your past work first. The model of you that VibeFlow draws from is no longer "the seven campaigns you've generated here," it's "the writing you've actually published over years." Your voice arrives the moment you arrive.

It also closes a quieter gap I've been bothered by since Phase 2 shipped: drift. Most users edit our output before publishing. They take what we generate, polish it, ship the polished version. We never see the polish. So our model of you slowly drifts away from the version of you that's actually in the world. The Voice Library lets you feed that real published version back in. Now we learn from what you ship, not just what you generate.

The through-line

A week ago, the Learning Engine made VibeFlow remember your voice across generations.

Yesterday, Vibe Tweak made your voice editable inside one generation.

Today, the Voice Library makes your voice the starting point — before generation one.

Past, present, future. Three different time-windows on the same idea: a marketing system that genuinely belongs to you only works when it's anchored in your taste, not the average taste, not the generic SaaS template. Tools that start over every Monday belong to OpenAI's distribution. Tools that compound belong to you.

What's coming next

This is Phase 2.5A in our roadmap. The follow-on is sequenced for the next four weeks:

Phase 2.5A v2 (Week 1). Add .docx and .pdf parsers, plus a proper Voice Library page with edit, search, and tags. Today's version is paste-and-URL only — adequate for bootstrap, not adequate for users with hundreds of pieces. The v2 work makes the library scale.

Phase 2.5B — Skills (Weeks 2–3). This is the bigger lift. Today, the Voice Library is implicit signal — Claude reads your past work and infers your voice. Skills make it explicit: VibeFlow analyzes your library and generates a structured voice spec for you (tone descriptors, lexicon, banned phrases, structural habits) that you can edit, lock, and apply on every generation. It's the difference between the AI gets your vibe and the AI knows the rules you write by.

Phase 3 — cross-user pattern detection (Weeks 2–3, in parallel). Anonymized clustering across all VibeFlow users: which X-thread structures actually drive replies, which subject lines actually open, which Reddit framings tank. Needs hundreds of opted-in campaigns plus engagement signals first — can't ship until there's data to cluster.

Phase 4 — Weekly Insights email (Weeks 3–4). Personalized summary built on Phase 3 clusters: "Your LinkedIn posts under 1,200 characters got 2.3× more views this week. Here's a tighter rewrite of last week's post."

Phase 5 — cross-agent coordination + public tactic library (Month 2+). When the Social agent writes an X thread, the Email agent knows about it and writes a follow-up that references it. All nine agents reading from the same playbook instead of producing disconnected assets. Plus a public, anonymized winning-pattern library as community knowledge.

The honest part

Voice Library is shipping before public launch — same posture as Vibe Tweak, same posture as Phase 2. The hypothesis is that closing the cold-start valley lifts first-week retention by ~10 percentage points. That's unproven. Thirty days of real usage will tell us whether founders actually upload their past work, or whether the feature reads as friction at signup.

If they don't upload, I'll know something needs to change about how we surface this. If they do, the marketing OS that compounds value the longer you use it just got a much faster ramp.

— Liz

---

VibeFlow Marketing is in pre-launch with friendlies. If you're an indie founder running your own marketing and want early access, drop us a note — we're letting people in as the system has data to learn from.

Found this useful? Share it.