Know you. Ground you. Protect you.

What we know about how you write, from your archetype to your sentence rhythm.

What we give the model to think against — your past hits, your goal, your strategy.

What we do to catch AI drift — at generation, at validation, at retry.

Archetype

A categorical voice positioning assigned at onboarding (e.g. Thoughtful Operator, Insight Builder, Story-First Connector) with its own reference sample post as a style anchor.

Tone calibration

Four sliders the user sets at onboarding and tunes over time: warmth, storytelling, provocation, openness. Fed in as numeric values with explicit interpretation guidance.

Vocabulary favors and avoids

Two paired lists of words and phrases the user reaches for naturally and words they actively avoid. Pulled from onboarding picks and learning-boost answers, refined over time.

Signature phrases

Up to three phrases the user identified as ones they reach for naturally. The model is instructed to use at most one per draft, only when it lands.

Stylistic anchors

Specific snippets from sample posts the user picked at onboarding as 'this is the hook style I identify with' or 'this is how I open.' Three concrete passages used as style anchors.

LLM-distilled voice signature

A multi-paragraph signature written by Claude from the user's actual published posts. Updated as new posts sync. Captures lens, observed patterns, and what makes the voice recognizable.

Signature moves

Six discrete recurring moves the user reliably reaches for, extracted from their post history. Things like 'opens with a specific stage moment' or 'ends on a hard recommendation.'

Stylometric fingerprint

Not just sentence length. We measure the unconscious habits that actually identify a writer: your punctuation profile (exclamations, semicolons, parentheticals, ellipses), paragraph and post length, how often you contract words, the openers and phrasings you reach for, your emoji/hashtag/list behavior, how you tend to open and close, and your I/you/we balance. Computed deterministically from your real posts — counts don't hallucinate — and fed in as hard numeric anchors.

Negative space

What you never do is as identifying as what you do. If you never ask questions, never use emoji, never reach for an em-dash, the model is told so explicitly. Most tools only imitate the patterns that are present; we also imitate the ones that are conspicuously absent.

Implicit learned preferences

When you edit a draft, an async pass extracts preference signals from what you changed. Stored, ranked by confidence, fed into the next draft. Captures voice instincts you might not even know you have.

Explicit learned preferences

When you complete a learning boost (an A/B 'which of these sounds more like you' prompt), your pick is stored with higher confidence than implicit signals. Conscious choices about voice carry more weight than inferred ones.

Topic-matched champion

Your best-performing past post on a topic similar to what you're drafting, surfaced live as the strongest single imitation target. Updates per workshop session.

Historical exemplars

Your top performers and your most recent posts, fed to the model as complete posts — never excerpts — with explicit 'match the cadence and rhythm of these' instruction. Whole posts keep your endings, your CTA, and your true post length intact; truncating them was quietly erasing the most identifying parts. Read alongside the voice signals as the imitation set.

Performance signals

Which post shapes, hooks, and CTAs actually work for you (and which the model might reach for as defaults). Acts as tiebreaker when multiple voice-aligned options exist.

Goal context

Whether you're trying to sound like yourself, get reach, start conversation, build authority, share a personal story, challenge a belief, or teach. Each shifts how the model writes.

Strategy context

Your active content pillars and perspective angles, so drafts stay on-strategy rather than drifting topic-by-topic.

Brand context

For managed brand profiles: mission, tone notes, taboos. Layered on top of the personal voice signals so brand drafts read on-brand without losing voice.

Conversation memory

Every prior turn in the workshop session, including the score rationales and your edit history, folded into the prompt. 'Apply your fixes' refers to something specific.

Hard content rules

Active 'don't write this' instructions: no em-dashes, no stacked single-sentence rhythm, no LinkedIn-influencer cliches, no tricolons or rhetorical-question closes as defaults. But every one of these defers to your real writing — if your own posts genuinely use em-dashes, we keep them, because banning them would erase your voice rather than the AI's. The rules suppress the model's defaults, not your actual habits.

User intent override

If you explicitly ask for one of the banned-by-default patterns (e.g. 'end with a question'), the system follows your request and doesn't second-guess you. The rules apply to defaults, not to your conscious choices.

Mechanical remediation

Post-generation: strips em-dashes, fixes paragraph structure, joins stacked single sentences, removes the obvious tells. Deterministic, fast, runs on every draft.

AI-structure validators

Detectors for tricolon overuse, rhetorical-question close, one-line-then-colon opener. Flag the draft and feed into the retry decision.

Authorship verification plus retry

Before you ever see a draft, a separate model judges it the way a forensic linguist would: could this pass as the same author as your real posts? It names exactly what gives it away (e.g. 'closes on a rhetorical question; you close on a flat statement in 4 of 4 samples') — and if the draft fails, the system regenerates once using those specific notes and keeps the better attempt.

Accuracy check

A second Haiku pass flags self-referential claims (e.g. 'in seven words' after a quote that's actually nine words) the deterministic checks can't catch. Trust-killers caught before publish.

Two feedback channels into one voice

Every edit you make becomes implicit preference data. Every learning-boost A/B pick becomes explicit preference data. Both flow into the same learned-preferences layer, ranked by confidence, fed into the next draft.

Workshop stance system

Multi-turn collaboration with three explicit stances (clarify, draft, refine). The model behaves like a thinking partner across a conversation, not a one-shot generator.

Compounding voice profile

Every published post adds to the voice signature, the signature moves catalog, the stylometric fingerprint, and the champion library. The longer you use it, the better the inputs.