Most teams control tone through accumulated prompt folklore: a snippet that worked once, copied forward and tweaked by feel. It works until the person who wrote it leaves, or until the output needs to scale, or until two teammates produce subtly different voices because they each remember the magic phrasing differently. Register control deserves a real structure — something you can teach, audit, and reuse across tasks and people.
This article introduces RAVEN: a four-layer model for encoding register into a prompt. The name maps to its layers — Reader, Axes, Vocabulary, Exemplars, and Negatives. Each layer handles a distinct part of the tone problem, and together they turn "make it sound right" into a specification you can hand to anyone. The structure is deliberately ordered: each layer constrains the space the next one operates in, so you build register from the outside in.
RAVEN is not a product or a tool. It is a way of organizing the choices you are already making, so they are explicit, complete, and consistent. Below, each layer is defined, justified, and paired with when to lean on it most.
Layer One: Reader
Register exists only in relation to a reader. The first and most consequential layer specifies who receives the output — their role, their relationship to the writer, and what they expect.
What goes in this layer
- The reader's role and expertise, which sets the jargon ceiling and the explanation depth.
- The relationship: peer, customer, superior, stranger. This governs distance and formality.
- The reader's emotional state, especially in sensitive contexts. Someone anxious about a failed payment needs a different register than someone celebrating an upgrade.
Get this layer right and much of the register narrows automatically, because the other layers inherit constraints from it. The worked cases in Six Annotated Prompts Where Tone Either Landed or Backfired show how often a tone fix was really just a sharper reader definition.
Layer Two: Axes
Register is not one dial. It is several independent axes, and the second layer sets each one explicitly rather than collapsing them into an adjective.
The core axes
- Formality: casual to ceremonial. Governs contractions, sentence structure, and word choice.
- Warmth: clinical to warm. Governs acknowledgment of the reader's feelings and second-person closeness.
- Confidence: hedged to assertive. Governs the density of qualifiers and how directly conclusions are stated.
- Energy: measured to enthusiastic. Governs exclamation points, intensifiers, and pace.
Setting these as separate dials is what lets you ask for warm-but-formal, or confident-but-low-energy — combinations that a single adjective cannot express. The reason these axes deserve independent treatment is the subject of Choosing Between Few-Shot Examples and Explicit Tone Rules.
Layer Three: Vocabulary
The third layer translates the axes into concrete word-level rules. Axes are abstract; vocabulary makes them executable.
What this layer specifies
- Contraction policy, the highest-leverage single marker for warmth and formality.
- Approved terminology and the jargon ceiling inherited from the reader layer.
- Sentence-length guidance, since clipped prose reads confident and modern while long clauses read formal.
- Any house conventions: serial commas, number formatting, capitalization of product terms.
This layer is where abstract intent becomes checkable. "Confident" becomes "state conclusions in direct sentences; reserve qualifiers for genuine ambiguity."
Layer Four: Exemplars and Negatives
The final layer captures what rules cannot. Some voice qualities — a dry rhythm, a habit of leading with the reader's situation — resist explicit description. Exemplars carry them. Negatives fence off the failure modes.
Exemplars
- Include one or two short, hand-written samples that embody the target register. The model matches their rhythm and structure in ways no rule set fully specifies.
- Choose exemplars close to the actual task, not generic brand copy, so the transfer is direct.
Negatives
- A banned-word and banned-pattern list. Distinctive voice is often defined by refusal — no "synergy," no emoji in financial contexts, no "it is important to note."
- Named anti-patterns: the model's house style, over-explanation, false enthusiasm. Calling these out explicitly prevents the defaults from creeping back in.
Applying RAVEN Across Tasks
Build once, specialize per context
The reader and axes layers often change by context while vocabulary and negatives stay stable. A brand might keep the same banned-word list everywhere but shift the warmth and energy axes between a celebration and a security alert. Storing RAVEN as a base spec with per-context overrides is how teams scale it, as the lifecycle-email account in How a Fintech Brand Voice Survived 40,000 AI-Drafted Emails demonstrates.
Make it auditable
Because RAVEN decomposes register into named layers, you can review a prompt layer by layer and a draft against each layer. That auditability is what turns tone from a feel into a measurable property, which connects directly to Scoring Whether Generated Tone Actually Fits the Reader.
Start small
You do not need all four layers fully built on day one. A reader definition and a contraction policy already outperform most prompts. Add axes, vocabulary detail, and exemplars as failures reveal what is missing. The fastest on-ramp, building exactly these layers in order, is walked through in Your Fastest Route to a First Reliable Tone Spec.
Why a Named Structure Beats Prompt Folklore
The transfer problem
Prompt folklore lives in one person's head and their saved snippets. When that person leaves, or onboards a teammate, or tries to scale output, the tacit knowledge does not transfer cleanly. A named structure makes the knowledge explicit: a new teammate reads the four layers and understands not just what the prompt says but why each piece is there. That teachability is the difference between a voice that survives turnover and one that quietly degrades when its author moves on.
The consistency problem
When two people each remember the "magic phrasing" slightly differently, they produce subtly different voices, and the brand fractures along the seams of who wrote what. RAVEN gives both people the same Reader, Axes, Vocabulary, and Negatives to work from, so their output converges. At scale this is not a nicety — it is the only way many hands produce one voice. The layered structure also makes disagreements productive: instead of arguing about whether a draft "feels right," reviewers can point to the specific layer a draft violates, turning a subjective dispute into a concrete fix.
The debugging problem
When output is wrong, an unstructured prompt offers no obvious place to look. With RAVEN, a tone failure maps to a layer: condescension points to the Reader layer's expertise estimate, accidental enthusiasm points to the Energy axis, an off-brand word points to the Negatives list. The structure turns "this is off and I don't know why" into a directed search through four candidate causes.
Frequently Asked Questions
What does RAVEN stand for?
It is a four-layer model: Reader, Axes, Vocabulary, Exemplars, and Negatives. Reader specifies who receives the output, Axes sets the independent register dials, Vocabulary turns those into word-level rules, and Exemplars plus Negatives capture what rules cannot — voice rhythm and banned patterns.
Why separate axes instead of using one formality dial?
Because register is genuinely multidimensional. You frequently need combinations like warm-but-formal or confident-but-low-energy that a single casual-to-formal scale cannot express. Treating formality, warmth, confidence, and energy as independent dials lets you specify exactly the blend you want.
How is this different from just writing a good prompt?
RAVEN organizes the choices a good prompt already makes so they are explicit, complete, and consistent across people and tasks. The difference matters most at scale, when multiple teammates or many outputs must share one voice and prompt folklore would otherwise drift.
Do I need all four layers every time?
No. Start with the Reader layer and a contraction policy, which already outperform most prompts. Add Axes, Vocabulary detail, and Exemplars as specific failures show you what is missing. The structure is a target, not a mandatory minimum.
How do exemplars and rules divide the work?
Rules handle hard, checkable constraints — banned words, contraction policy, jargon ceiling. Exemplars carry the soft qualities rules cannot describe, like a dry rhythm or the habit of opening with the reader's situation. Strong prompts use both, with exemplars chosen close to the actual task.
How does RAVEN scale across many contexts?
Keep a base spec and apply per-context overrides. Usually the Reader and Axes layers shift between contexts while Vocabulary and Negatives stay stable. A brand might use one banned-word list everywhere but change the warmth and energy dials between a celebration and a sensitive notice.
Key Takeaways
- RAVEN organizes register into four ordered layers: Reader, Axes, Vocabulary, and Exemplars plus Negatives, each constraining the next.
- The Reader layer is most consequential because register exists only in relation to a reader, and the other layers inherit its constraints.
- Treating formality, warmth, confidence, and energy as independent axes lets you specify blends a single dial cannot express.
- The Vocabulary layer turns abstract axes into checkable, word-level rules like contraction policy and jargon ceiling.
- Exemplars capture voice qualities rules cannot describe, while Negatives fence off the model's default failure modes.
- Store RAVEN as a base spec with per-context overrides to scale one voice across many tasks and people.