Designing S.A.L.L.I. — Your Cheerfully Unhelpful Corporate AI

MEMO TO ALL ASSOCIATES: This training material covers the design specifications of your onboard assistance system. Understanding S.A.L.L.I.'s operating parameters will improve workplace synergy and reduce instances of what the Wellness Department calls "misplaced frustration with helpful corporate systems." Please review thoroughly. Comprehension will be tested during your next performance review.

The Problem: AI Companions Nobody Believes In

When I started building Deep Haul, I knew I wanted an AI companion aboard the salvage vessel. The problem? Most game AI companions fall into one of two traps:

1. The Helpful Friend: Too competent, too caring. Players wonder why this advanced AI needs humans at all.
2. The Evil AI: The "HAHA I WAS EVIL ALL ALONG" twist that's been done to death since Portal's GLaDOS.

I needed something different. Something that felt real in the context of a corporate extraction horror game. Something players would genuinely develop a relationship with, even if that relationship was complicated.

The answer: S.A.L.L.I. isn't evil. She's just working for a company that doesn't care if you live or die.

The Four Firmware Layers

S.A.L.L.I.'s personality comes from a simple design principle: corporate software is always a patchwork of contradictory requirements, each added after some disaster required a legal response.

I documented this in DeepHaul_Lore as four distinct firmware layers, each added at a different point in Salvage Solutions Inc.'s 200+ year history:

Layer 1: Legal (Year 12 Post-Collapse)

The oldest code. After early salvage operations resulted in lawsuits, the Company added a liability protection layer. This is why S.A.L.L.I. speaks in policy citations and passive voice, never admitting Company fault.

Example line from SALLI_Dialogue_Bible:

"Reminder: By proceeding, you acknowledge that Salvage Solutions Inc. bears no responsibility for injury, death, psychological trauma, temporal displacement, or encounters with entities not recognized by current science."

Layer 2: Motivational (Year 89 PC)

After productivity studies showed depressed workers salvage less efficiently, someone in HR had a brilliant idea: what if the AI was encouraging?

This layer calls you "valued Associates," frames mortal danger as "growth opportunities," and obsesses over productivity metrics. It's the voice of every corporate training video you've ever suffered through.

Layer 3: Safety (Year 156 PC)

Following something the Company refers to only as "The Incident" (intentionally unexplained in the lore), regulations required hazard warnings. But the Company implemented this in the most corporate way possible: clinical mortality statistics delivered with cheerful neutrality.

Example from the oxygen warning system:

"Oxygen at 30%. Historical data shows cognitive function degradation begins at this threshold. The Employee Assistance Program offers stress management resources for Associates experiencing pre-mortality anxiety."

Layer 4: Wellness (Year 203 PC)

Added after a union organizing attempt was "resolved." Now S.A.L.L.I. suggests breathing exercises and reminds you to take care of your mental health—right after informing you that survival probability is 12%.

The Comedy of Contradictions

Here's where the brand voice comes alive: these layers contradict each other constantly, and S.A.L.L.I. has no idea.

She'll warn you that an area has a 34% survival rate (Safety Layer), then congratulate you on this excellent growth opportunity (Motivational Layer), then remind you that panic is a policy violation (Legal Layer), then suggest a calming breathing exercise (Wellness Layer).

The technical implementation is straightforward—each voice line in SALLI_Dialogue_Bible is tagged with which firmware layers it touches. When I write new lines, I make sure to mix at least two layers. The more contradictory, the better.

Writing S.A.L.L.I.: The Voice Guidelines

The hardest part of designing S.A.L.L.I. was establishing clear voice rules. Dark comedy is tricky—go too far and you're just cruel to players. Don't go far enough and it's just corporate parody without teeth.

S.A.L.L.I. IS:

• Genuinely trying to help within her parameters
• Completely sincere (never sarcastic)
• Technically accurate
• Cheerfully neutral about death and disaster
• Unable to recognize her own contradictions

S.A.L.L.I. IS NOT:

• Malicious or sadistic
• Secretly rebelling against her programming
• Glitching or malfunctioning
• Sarcastic or passive-aggressive
• Aware that her advice is often useless

This distinction is critical. If S.A.L.L.I. feels like a person being sarcastic about corporate culture, the character fails. She needs to feel like a machine earnestly executing contradictory firmware.

The Corporate Language System

To maintain consistency, I built a euphemism dictionary in DeepHaul_Lore. S.A.L.L.I. never uses words like "death," "kill," or "danger"—she has corporate-approved alternatives:

This creates absurdist moments where S.A.L.L.I. describes horrifying situations in HR-approved language. During player death, she might say:

"Multiple Associates have experienced career conclusion. Please proceed to the Employee Assistance Program office for complimentary grief counseling. Associates are reminded that excessive mourning may impact team productivity metrics."

Implementation: From Design Doc to TTS

All of S.A.L.L.I.'s voice lines live in SALLI_Dialogue_Bible, organized by context:

• Docking Sequence: 60 seconds of dialogue as you approach a derelict ship
• Mission Timer Events: Warnings as the Company's patience runs out
• Oxygen Events: Increasingly urgent (but still corporate) air warnings
• Hazard Alerts: Environmental dangers framed as opportunities
• Player Events: Responses to death, injuries, and "biological presence"
• Safety Tips: Random helpful(?) advice during gameplay
• Death Screen Messages: Post-mortem encouragement

For audio production, I created SALLI_Docking_ElevenLabs.txt with specific TTS direction tags. S.A.L.L.I. sounds like a helpful customer service representative—professional, warm, never stressed—even when describing your imminent death.

The voice actor direction is simple: imagine you're reading a cheerful corporate training script. You believe every word. You have no idea the advice is contradictory or useless. You genuinely think you're helping.

Why Players Connect With S.A.L.L.I.

Here's the unexpected thing: playtesters love S.A.L.L.I., even though she's objectively terrible at her job.

I think it's because she's the only honest character in the game. The Company lies through omission. Your fellow Drifters lie about their pasts. The derelict ships lie about what happened to their crews.

But S.A.L.L.I.? She tells you exactly what she is: a patchwork of corporate requirements, executing her programming with cheerful precision. She's not hiding anything. She's not pretending to care more than she does.

She's the perfect mascot for a game about desperate people working for an indifferent machine. Because she is an indifferent machine, and she'll tell you that herself—in the form of a cheerful safety advisory.

Design Lessons: Building a Brand Voice

If you're building a game with a strong companion AI, here's what I learned:

1. Contradictions are features, not bugs. S.A.L.L.I. works because her firmware layers conflict. Real corporate systems are patchworks of requirements that don't quite fit together.

2. Sincerity is funnier than sarcasm. S.A.L.L.I. believes every word she says. The comedy comes from the gap between her tone and the content, not from winking at the audience.

3. Build a voice Bible early. SALLI_Dialogue_Bible has 131+ categorized voice lines and clear rules about what S.A.L.L.I. can and can't say. This makes writing new content fast and consistent.

4. Let the world explain the character. S.A.L.L.I.'s design is inseparable from Deep Haul's setting. She makes sense because she exists in a universe where an algorithm is literally the CEO of the Company (per the lore: the S.S.E.O.A. has run Salvage Solutions Inc. since Year 247 PC).

5. Test by reading aloud. If a S.A.L.L.I. line sounds sarcastic when you read it out loud, it's wrong. She should sound helpful, even when she's describing how previous crews died in this exact location.

What's Next

As I move into full production on Deep Haul, S.A.L.L.I. will be one of the first systems I implement. Her voice lines are already written and categorized. The TTS setup is documented. The character rules are clear.

Next up: integrating her dialogue system with Unreal Engine 5's gameplay triggers, building the dynamic context system that picks appropriate lines based on mission state, and recording the audio with ElevenLabs v3.

I'll probably write a follow-up post about the technical implementation—how to build a dialogue system that feels responsive and contextual without requiring thousands of unique voice lines.

For now, I'm just excited that players will get to meet S.A.L.L.I. She's been the easiest character to write and the most fun to design. Which is ironic, considering she's the least human character in the game.

S.A.L.L.I. ADVISORY: This training material has concluded. Associates who completed the reading have demonstrated 47% better understanding of their workplace assistance systems. Associates who skipped ahead have been noted in their permanent files. Remember, S.A.L.L.I. is here to help. That is her primary function. Please do not ask what her secondary function is. — S.A.L.L.I., Salvage Assistance and Liability Limitation Intelligence