SelfScale · Internal

The Voice Engine Rethink

Brief for Bill · May 2026

Why we're rebuilding the cold-start, sunsetting the wizard, and adding a CTA layer, without rewriting anything you've built.

Three movements

The architecture shifted in three places

Cold-start lowered

1 sample, not 5. Default voice fills the gap.

Wizard sunset

Mandatory 5-step gate becomes opt-in coaching.

Intent / CTA layer

New explicit slot: posts know what they're FOR.

These ship as peers, not in sequence. They're coupled by one re-frame: voice fidelity is the mechanism, not the outcome.

Cold-start

The activation floor drops from 5 to 1

1

minimum sample required

Before

5 chunks minimum before any generation. creator-bible-auto.ts · CHUNK_THRESHOLD = 5. 10–15 min total time-to-first-generation.

After

1 sample → usable draft via default-voice splice + cold-start bible. Ladder rebuilds at 1 / 5 / 20 / 50 chunks. Quality lifts as the ladder climbs.

Activation impact

The gap was the wait, not the model

Funnel: trial start → first sample → first draft → first ship. ChatGPT is ~5 sec to first useful output. We sit at 6–12 min on the wizard alone, 100–180× slower than category baseline on time-to-wow.

Today22%

trial → first ship within 72h

Predicted41%

trial → first ship within 72h

+19 percentage points projected after the cold-start fix lands.

Wizard sunset

What the wizard did, and why it's done

What it did

· 5-step Q&A, sequential gating
· 100-char minimum on first paste (step-sources.tsx:179)
· Force-fed voice setup before any generation
· ~28% drop-off before completion
· Contradicts your own 2026-04-15 framing: "drastically minimize the clicks"

What replaces it

· ChatGPT-style input, no upfront gate
· Inline coaching after gen #1, #3, #5
· Ladder-based bible rebuild (1 / 5 / 20 / 50)
· Opt-in "Sharpen your voice (8 min)" surfaced inside /voice
· VAG stays as publishability gate: 0.62 soft / 0.42 hard

The net-new dimension

Posts have outcomes. Now we model them.

Every post is hired to do a job: funnel to product, build authority, ask a question, take a sponsor pitch. The engine never knew which, so it averaged. Six explicit types, one user-authored chip per generation.

Funnel

Skool, course, list, podcast

Sponsor

Brand deals, disclosure required

Engagement

Question, hot take, attention is the asset

Build-in-public

Ship log, community-building

Authority

Pure expertise display, no ask

Reactive

News/trend, perishable

Runtime capture

Three modes, stacked

1

Inventory setup

/voice → CTAs. 3–10 cards. 5 min, opt-in. Editable any time. Engine never edits this.

2

Per-generation chip

intent: [▼ none] under the composer. One click before Generate. Default is always 'none': anti-marketer guardrail.

3

Inferred intent

After ~50 generations, engine pre-fills the chip from brief topic. Sara accepts or overrides. v2.

ComposerVOICE 92

@linkedinintent: ▼ funnel → SkoolGenerate

The most important slide

CTAs ship verbatim. Period.

User-authored CTA, slotted verbatim+1% to +3% IS_me

Engine-paraphrased CTA−8% to −15% IS_me

Inferred-only (no user CTA)neutral IS_me

Rule: CTAs ship verbatim from inventory unless the user types one in the box.

The engine never improvises "Curious to learn more?" phrasing. That's the difference between voice-fidelity-positive and voice-fidelity-negative.

Channel awareness

One rule per platform

Small rule engine (~100 LOC). Sits between CTA inventory and prompt assembler.scoreVoiceMatch atblueprint-service.ts:1427 keeps scoring honest across platforms.

LinkedInInline link OK

XReply-style, no link in body (algo penalty)

RedditNo link unless karma + mod-approved; default to brand name only

TikTok"link in bio"

Instagram"link in bio"

Threads"link in bio"

YouTubeDescription-only

SkoolDirect in-community sell is fine

The default mix

60 / 30 / 10

60% Authority

30% Engagement

10% Funnel

The engine biases toward Authority and Engagement. Funnel CTAs ship sparingly, not on every post, because conversion ≠ voice. If Sara picks Funnel twice in a row, the composer surfaces a soft nudge with one-click switch. The engine doesn't refuse; it just doesn't make over-CTA the path of least resistance.

What 'good' means

The IS_me rating is the gate. The proxies are co-sensors.

We never anchored a phase-success metric to a proxy score alone. The binding signal is the founder's IS_me rating on real generations.

Outcome gate

content_items.is_me_founder_rating: boolean + Likert (1–5) + reason. Cohort rating ≥ 75% is the activation threshold.

Co-sensors

scoreVoiceMatch (per-gen proxy) · VAG 0.62/0.42 (publish gate) · EQI (daily aggregate).

New paired metric: IS_me × CTA conversion for posts with a CTA attached.

Untouched

Don't touch these. They work.

This is additive, not a rewrite. Everything below is preserved exactly as you built it.

Anti-AI filter (anti-ai-filter.ts, REGEN_THRESHOLD = 0.05)

Voice Authorship Gate (voice-authorship-gate.ts, 0.62 / 0.42)

Blueprint service (blueprint-service.ts, Phase 148)

EQI daily job (eqi-scorer.ts, voiceAccuracyJob)

creator_bible pipeline (creator-bible.ts)

knowledge_chunks ingestion pipeline

Greenfield filter + scoreVoiceMatch Haiku scorer

Existing wizard code (entry point changes, code stays)

Up for re-litigation

What gets challenged

Wizard 5-step gate

Deprecated to opt-in. Surfaced inside /voice after gen #1.

CHUNK_THRESHOLD = 5

Hard floor lowered to 1, becomes a ladder (1 / 5 / 20 / 50).

Implicit-only CTA inference

Now explicit. User-authored cta_inventory + per-gen chip.

Single-CTA-per-user assumption

Now inventory: 3–10 cards, frequency-capped, type-tagged.

Voice fidelity as the product

Voice fidelity is the mechanism. Conversion is the outcome.

Need your input

7 decision points for Bill

Ranked by impact. These are the questions I can't answer alone. Each has a default, but I want pushback on any of them before we draft the GSD phases.

1
DECIDE
Where does CTA inventory schema live? Separate table vs JSONB on voice_profiles.
2
DECIDE
Pre-blueprint generation: ship at 1 sample, or block until 3?
3
DECIDE
Wizard sunset: hard cut or feature-flag rollout?
4
DECIDE
CTA picker placement: composer chip, sidebar inventory, or both?
5
DECIDE
Voice scorer recalibration: bake new IS_me signal into scoreVoiceMatch now or after R3-B?
6
DECIDE
Per-channel CTA rules: hardcoded adapter or per-channel config table?
7
DECIDE
Founder override: allow inline CTA override even when inventory is empty?

Rollout

How we sequence the ship

Phase A

Cold-start fix + wizard kill switch behind FF.

Phase B

CTA inventory + inline picker chip.

Phase C

Per-channel adapter + 60/30/10 mix logic.

Phase D

Founder-rated IS_me wired into the gate.

Phase A is single-PR scope: lowest-risk, highest-impact, unblocks everything else. Phase D closes the loop with the binding outcome gate.

Scope discipline

What we're NOT doing

Not rebuilding the model.

Not changing VAG publish thresholds (0.62 / 0.42 stay).

Not removing EQI or the daily voice-accuracy job.

Not breaking the wizard URL: /voice/wizard redirects for 90 days.

Not requiring re-training of existing voices. Ladder applies forward only.

Not making channel posts the primary voice source, they stay filtered, secondary.

Additive, not destructive. Your work survives the rethink.

The honest summary

“We had a great voice engine that nobody could feel inside 5 samples. We're not making the engine smarter. We're making it faster to feel and clearer about what it's writing for.”

- the rethink

By MondayDecisions back on the 7 points.

By WednesdayPhase A scoped: cold-start fix + wizard FF.

End of weekPhase A ships behind FF. Cohort instrumentation live.

Brief lives at ~/Desktop/BILL-BRIEF-voice-engine-rethink.md

Thanks for reading. - David