AI Engineer World's Fair 2026

Stop Writing Tone Instructions. Layer Them.

Loading slides…

Read the talk

Stop Writing Tone Instructions. Layer Them.

A wedding-venue assistant needs more than a warm voice: ordered identity rules, situational context, examples, and an output gate that can refuse to send an unsupported promise.

From a talk by Isadora Martin-Dye

Before you start: Familiarity with system prompts and the distinction between model generation and application-side validation will help; the code examples use TypeScript.

When the examples run out

What happens when a wedding-venue assistant gets a question its brand examples never anticipated? Isadora Martin-Dye approaches that problem as a venue owner: she runs a Virginia wedding venue in a 225-year-old property and built an assistant for her couples, then assistants for other venues, a personal companion app, and a utility for families of missing people. Her working analogy is a brilliant intern with high intelligence and poor emotional judgment. The intern remembers the induction material but cannot reliably read the room. Managing that intern requires structure and checking their work before it leaves the business.

A detailed system prompt can describe the brand, supply phrases, and demonstrate good answers. That works on the happy path. Then comes Martin-Dye’s illustrative turn twenty-one: the first question the examples do not cover. The answer may be technically correct yet sound like something the business would never say.

Slide titled “It works for the happy path,” with text about turn 21 and a code comment reading “Write in our brand voice.”Open full source frame
The happy path ends at the first question the examples did not anticipate.

That mismatch matters especially when the relationship is part of the product. A luxury hotel with decades of cultivated service, a high-end real estate firm, or a wedding venue cannot treat voice as decoration around a factual answer. Customers are paying for a relationship and notice when it breaks. The underlying problem is that an instruction to write in the brand’s voice asks one prompt to perform four different jobs.

0:000:16
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:00 · section reference included

One assembler, four responsibilities

The architecture separates those jobs into four layers:

LayerResponsibility
Immutable identityDefine constraints that customization must not override.
Situational modeAdapt to the audience and present circumstances.
Example-anchored voiceSupply warmth, phrases, examples, and tone settings.
Post-generation vetoInspect the actual response and catch violations.

Most brand-prompt work concentrates on the middle: how to sound and, sometimes, whom to address. It leaves the boundaries and the final inspection underdeveloped.

Martin-Dye reports replacing twenty-four scattered system prompts with one canonical assembler. Some prompts called the assistant Sage, others Venue, and others supplied no name at all; each surface had developed its own identity. Every narrator now uses the same entry point to compose its prompt. The assembly order matters: hard rules first, task last. The fourth layer operates after generation rather than being another paragraph in that prompt.

Slide titled “One assembler. Fixed order.” showing a systemPrompt array of rule, voice, context, guard and task blocks.Open full source frame
One assembler composes the prompt in a fixed order: hard rules first, task last.

Google Maps provides the organizing analogy. The destination stays fixed, but traffic, roadworks, fuel needs, and route preferences affect how to get there. A license requirement and a prohibition on driving backward down a motorway are constraints; traffic is a condition; a scenic route is a preference. A final route check happens before departure. The assistant likewise needs the relevant conditions before producing its answer, assembled in a fixed order rather than discovered after it has taken the wrong turn.

2:122:27
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

2:12 · section reference included

Identity sets the boundaries of warmth

Immutable identity defines policy precedence. These are constraints, not preferences that a venue’s personality settings may replace. If someone asks whether the assistant is human, a bot, or a live agent, the policy requires an unambiguous confirmation that it is an AI assistant in the very next message. Bloom also discloses that identity in its first response, before anyone asks. Venue configuration, voice profiles, personas, and user requests do not override that policy.

Martin-Dye presents upfront disclosure as a product decision about trust: someone who knows from the beginning can understand the relationship they are entering, rather than discovering later that they misunderstood it. The rule belongs above the venue’s tone choices.

The physical-presence boundary makes the distinction concrete. Software has no body: it cannot meet a couple or personally conduct a tour.

LanguagePolicy
“I'd love to show you around”Forbidden: implies physical presence.
“I can't wait to meet you in person”Forbidden: promises an impossible meeting.
“The team would love to host you for a tour.”Allowed: assigns the visit to the team.

The voice layer tends toward warm first-person language. Without a boundary, that warmth can produce a lie about who will be there. Martin-Dye describes the resulting discovery as an inversion of trust: the user realizes they have been building a relationship with someone who was never physically present.

4:525:07
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:52 · section reference included

When a natural word carries unjustified certainty

Martin-Dye applies the same architecture to Threadline, her tool for families of missing people, with a very different voice and much higher stakes. In her account, its identity rules prohibit certainty-laden words such as confirmed, identified, matched, proven, linked, and solved. A probabilistic indication must not become a statement that someone’s missing loved one has been found.

A model may reach for the word match because it is a statistically natural continuation. A parent who has spent years wondering where their child is can hear something much more definitive. The rule therefore governs what the product may claim, not simply whether its wording sounds compassionate. Different products need different prohibitions, but the structural requirement stays the same: voice customization must not grant permission to make claims the product cannot support.

6:587:09
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:58 · section reference included

Change the route for the person

Situational mode supplies real-time conditions before generation. Sending the same system prompt to everyone ignores information that should change the response. Just as a navigation system incorporates an accident or a driver’s fuel needs before choosing a route, the assistant needs to know who is asking and what conditions apply.

The same assistant speaks to couples and briefs venue staff. With a coordinator, it should speak as a colleague rather than a customer-service representative, while retaining its character instead of switching into generic intelligence-analysis prose. If a coordinator asks whether inquiries will rise in June, an appropriate answer admits that it cannot forecast confidently and then explains the trend and what the coordinator might see.

Martin-Dye distinguishes that staff-facing response from how the assistant should address a couple. The identity and voice remain consistent, but the audience changes the response strategy. The next signal goes beyond someone’s role to their life: what is this particular person going through?

8:248:38
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:24 · section reference included

Human context before numbers

Soft context notes guide tone, empathy, and what not to say; the assistant must not quote them verbatim. A mention of grief should produce gentleness, not a recital of the loss. A sick parent should lead to patience and flexibility about timing, not an unsolicited sentence about the illness.

The assembler deliberately renders the couple’s notes before the numbers-guard instructions. Human context establishes the tone, and numeric constraints follow. Martin-Dye attributes mechanically slotted prose to reversing that order: the model commits to a numeric framing before receiving the qualitative context. This ordering can be expressed directly in a TypeScript assembler fragment:

typescript

function assembleContext(
  coupleNotesBlock: string,
  numbersGuardBlock: string,
  taskBlock: string,
): string {
  return [coupleNotesBlock, numbersGuardBlock, taskBlock]
    .filter((block) => block.trim().length > 0)
    .join("\n\n");
}

Here numbersGuardBlock contains instructions supplied before generation; it is distinct from the later check that can reject generated output.

The concrete example is a heat map of contact frequency. A couple’s declining engagement means something different when the system knows the client’s mother has been in chemotherapy for three weeks. The numeric movement is the same, but the human context changes its interpretation.

Slide titled “Driving past the roadworks doesn't make them go away,” with comments about a heat drop during a parent's chemotherapy and rendering couple notes before the numbers guard.Open full source frame
Human context sets tone before numeric constraints.

Layer one prevents the assistant from pretending to be something it is not. Layer two lets what it actually knows shape its behavior. With that context, falling engagement reads as a family under strain, rather than a cold lead or a difficult couple to chase. The voice need not change for the response to become more appropriate.

10:0110:15
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:01 · section reference included

Examples teach expression, not guarantees

The third layer is the familiar tone guide: dials, phrase lists, and examples of good responses. It often arrives from marketing, gets wired into the prompt, and marks the end of the engineering work. In the intern analogy, it is the induction pack handed over on the first morning. Useful as that pack is, it was prepared before today’s customer walked through the door.

Examples cannot take over the other layers’ responsibilities. They do not enforce an identity boundary, supply the current person’s circumstances, or inspect what the model has just written. When an unfamiliar question arrives, a phrase list has no guarantee to offer. That is a mismatch between the tool and the job, rather than evidence that examples are useless.

Bloom’s voice-training workflow lets venues adjust their brand voice and incorporates coordinators’ ratings of actual AI responses and their edits. That makes the examples responsive to real work. Martin-Dye describes learning from those edits without specifying whether the implementation changes prompts, stored examples, or model weights.

11:4511:58
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

11:45 · section reference included

Read what actually came out

The post-generation veto is the only layer that reads the completed response. The first three layers instruct the model; this one inspects the result and can stop it from leaving the business. It fills the role of the person who reviews a new intern’s client email before it is sent.

Martin-Dye separates two kinds of intervention:

  • Soft flag: The honesty inspector flags a response that may have slipped a boundary, including concern about whether it answered the question. A false positive means someone double-checks an acceptable response. A false negative can let a hallucinated number or privacy violation reach a client.
  • Hard reject: The numbers guard rejects output containing a figure the model was never given, even if the prompt already told it not to invent numbers.

The distinction is operational: a flag asks for inspection; a rejection denies permission to send. The asymmetric consequences of missed errors are why this work needs an explicit place in the system.

13:1713:28
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

13:17 · section reference included

A beautiful promise for a booked Saturday

The veto grew out of a recurring failure. A couple would ask about a Saturday in October. The assistant would respond warmly about the property at that time of year and offer to hold the date—except the date was booked. The model had never received the calendar. It knew the shape of an encouraging service response without knowing whether the promise was possible.

The identity rules could hold, the audience mode could be correct, and the voice could be perfect while the response still failed. Indeed, confidence and warmth made the unsupported offer more persuasive. The couple now believed they had a date; the business had merely postponed their disappointment.

This exposes the limit of the word immutable: policy precedence does not guarantee model compliance. All three prompt layers remain probabilistic instructions. That may be tolerable for a slightly off-brand sentence, but a date, price, policy, or promise is something a person may act on. Martin-Dye calls the veto the cheapest layer to build: it checks specific generated claims against permitted facts. A date offered outside the allow list does not ship. The prompt provides prevention; the gate provides enforcement for the checks it implements. Both are needed because an instruction can eventually lose to generation behavior.

14:3114:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:31 · section reference included

Missing tenant identity is an error

The integration seam is one function. Across venue tenants, layer one stays shared while layers two and three load venue-specific conditions and voice. The same root architecture supports different venue personalities and products such as Ground and Threadline without forking the underlying logic. Sharing that logic, however, creates a dangerous temptation to share identity defaults.

Martin-Dye describes a white-label leak in which silent defaults caused venues to ship as Sage using another venue’s email address. Missing configuration had become another venue’s identity. The fix was to require brand identity from the venue AI configuration and throw an error when it was absent. In the navigation analogy, every driver had been receiving the same saved home address.

Tenant identity must fail loudly rather than fall back. The essential validation is small:

typescript

type VenueAIConfig = {
  aiName?: string;
};

function requireAIName(config: VenueAIConfig): string {
  const name = config.aiName?.trim();
  if (!name) {
    throw new Error("Venue AI configuration requires an AI name");
  }
  return name;
}

A missing name stops the path instead of silently substituting Sage. That visible failure is preferable to a venue speaking in a stranger’s voice while the team remains unaware that anything is wrong.

Slide titled “One architecture. Different drivers.” showing a function that checks the configured AI name and throws an error if it is missing, plus a warning against default brand identity.Open full source frame
Missing brand identity must raise an error rather than use a default.
16:2716:46
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

16:27 · section reference included

Make permission unavoidable

“The first three layers are instruction, the fourth is permission.” Identity, conditions, and voice tell the model what to do. The output gate decides whether the response may leave the business. Separating constraints, situational judgment, expression, and inspection turns a single overburdened prompt into distinct engineering responsibilities.

Customers are trusting the voice, not testing it. Martin-Dye’s four-layer pattern comes from repeated product failures rather than a framework designed in advance. The practical question is whether the next failure will be discovered during testing or by a customer. That also requires examining where the architecture is still easy to bypass.

At the time of the talk, the numbers guard lives inside a heat-narration path, while the honesty inspector is a separate function. Each new surface must remember to wire in the veto manually. Martin-Dye proposes making it a shared service: a mandatory gate that every surface passes through by default. This is a proposed improvement, not an already completed property of the system she describes.

Condition selection has a similar gap. Layer-two mode detection remains partly manual: the heat-narration surface knows to load couple notes, while the briefing surface knows not to. A central condition resolver would make those choices explicit in one place, determining who the user is, what circumstances matter, and which route applies.

The final trade-off concerns the soft flag itself. It uses regular expressions, not a model. Regex provides fast, inexpensive, repeatable matching for encoded patterns; it does not establish complete semantic coverage. A small classifier might catch edge cases no one anticipated when writing those patterns, but would introduce probabilistic classification errors. Martin-Dye currently chooses determinism over broader coverage and would make that choice again. She leaves it as a real engineering trade-off, not an obvious win: predictable enforcement and comprehensive detection are different properties.

17:4717:59
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

17:47 · section reference included

Resources

From the talk

  • Overview of Bloom's venue assistant, voice feedback, couple portal, and multi-venue branding, with a link to a sample-venue demo.

Transcript download
Full transcript

Loading transcript…