The Human-Centric Sandwich

The AI-assisted writing system Aureum uses for its own content production — and demonstrates to clients. Four layers: human input → AI synthesis → AI refinement → human truth-telling. The AI does the structural work. The humans do the nuance. This is the methodology that lets the Aureum content engine produce 45–55 pieces per month without burning out a two-person studio.

The four layers

Layer 1 — Input (Human, ~10 minutes)

Record a 5–10 minute voice memo or write a bulleted outline covering:

  • The core insight or observation — what's true and why it matters
  • The specific evidence or project context — the numbers, the client moment, the concrete detail
  • What you want the reader to think or do differently afterwards
  • The tone — more analytical or more personal

Format freedom is intentional — messy voice notes work as well as clean outlines. The human layer is about substance capture, not polish.

Layer 2 — Synthesis (AI, Automated)

The AI expands the input into a structured first draft using the Aureum brand voice guide.

System setup: - Claude Project or custom GPT trained on a corpus of Aureum's best writing — Integration Files, Design Briefs, Aureum Notes - Plus the Brand Voice Guide as a grounding document - Plus relevant passages from the monthly topic bank

The synthesis layer produces a complete first draft — structurally sound, voice-calibrated, ready for refinement.

Layer 3 — Refinement (AI, Automated)

A second AI pass checks the draft against Aureum's Studio Values brief. Flags anything that needs human decision.

Refinement checks: - Is every claim backed by evidence or experience? - Is there a single clear CTA (not three)? - Does it sound like Auria — or like generic AI? - Are any forbidden phrases present? - Does it respect British English throughout? - Is the piece one idea clearly argued, or many ideas diluted?

Flags go back to the human layer. Passes proceed to final edit.

Layer 4 — Final Edit (Human, ~15–20 minutes)

Add the specific detail only the humans know. This is not polish — it's truth-telling.

  • The actual number from the project
  • The thing the client said
  • The decision that felt hard
  • The nuance that makes it real

This layer is the difference between AI content and Aureum content.

Why it works

Most AI-generated content is generic because it starts with a generic prompt and ends with generic output. The Sandwich works because:

  1. Specificity enters at Layer 1 — the human input carries real experience, real stakes, real detail
  2. Voice is a system, not a prompt — the AI is trained on actual Aureum writing, not instructed via personality descriptors
  3. Refinement is a separate pass — catches failures the generation step can't see
  4. Truth arrives at Layer 4 — the AI scaffolds the shape, humans inject the soul

The AI extract is a scaffold. The voice and authority are the founders'.

Time economy

Layer Time Who
Input ~10 min Human
Synthesis ~2 min (automated) AI
Refinement ~2 min (automated) AI
Final Edit ~15–20 min Human
Total per piece ~25–30 min human time

This makes 45–55 pieces per month possible. Without the Sandwich, that volume would require full-time content writing. With it, both founders contribute ~1–2 hours/week in Month 1, dropping to 30 minutes as voice consistency compounds.

Training the system

The Sandwich only works if the AI has been properly trained. Training inputs:

  1. Corpus of Aureum's best writing — 15–25 exemplar pieces across all three Substack series
  2. Brand Voice Guide — what Auria does and never does
  3. Forbidden Phrases — banned words and patterns
  4. Auria vs María-Inés — voice-routing rules
  5. Content Pillars — the strategic territories every piece must sit within
  6. Past client case notes (anonymised) — specific detail patterns, numbers, framings

The first 4–6 months of content requires active founder review at Layer 4. After that, the voice becomes self-reinforcing — the AI has enough corpus to stay inside the register without drift.

The demonstration effect

Aureum doesn't just use the Sandwich internally — the methodology itself is a demonstration of what Aureum builds for clients.

The same system we use to scale our content is the same system we build for founders who have substantial IP but cannot scale their voice fast enough. Content is the proof of concept.

When a prospect asks "how do you make AI sound like me and not like AI?" — the answer is the Sandwich. Shown, not told.

Sources

  • sources/raw/substack-strategy.md — The AI Content System, First Draft Architecture section
  • sources/raw/content-engine-full-strategy.md — How the Sandwich feeds the monthly engine