#conversation design

This month’s interview is with Greg Bennett, a Director of Conversation Design at Salesforce, a customizable, cloud software platform for customer relationship management. In this interview, Greg discusses the culture shock from transitioning to industry from academia, and how his discourse analysis training has impacted his work and career in the tech industry. You can find Greg on Twitter @gabennett45 or on LinkedIn.

This last post of 2022 is also the last in this series, and we'll have more in 2023 about the future of lingjobs!

What did you study at university?

I hold a BA in Linguistics and MS in Applied Linguistics from Georgetown University. I also studied Japanese language and culture during undergrad. My academic focus was on interactional sociolinguistics—particularly, how users manipulate textual stylistics to convey contextualization cues during synchronous, text-based chat.

What is your job?

I founded and lead the Conversation Design team and practice at Salesforce. I oversee strategy and workflow for an international team of Conversation Designers who craft conversational experiences that “sound” like Salesforce and train language models to recognize myriad varieties of language input from users. I spend much of my time discussing with various leaders across the business the resource investment strategy for conversational apps and features that position Salesforce at the forefront of the market; scaling our resources and tools to expand my team’s sphere of influence across our ever-growing lines of business; and championing my team’s impact on our bottom line within and beyond the company.

How does your linguistics training help you in your job? 

I never expected linguistics to be as vital or central to my role as it is now. I’ve always had my discourse analytic mindset running in the back of my mind to interpret conversation and stancetaking in real time in all of my past roles. As a leader at Salesforce, I certainly maintain that approach to developing relationships with stakeholders across the business. However, since conversational AI is fundamentally about language as an interface for a technological system, I find myself referencing every aspect of my training as a linguist to contextualize and strengthen my proposals for a conversational solution—from generative syntax when debating the ‘conversationality' of a conversational app’s brand name, to acoustic phonetics when determining the pitch range for a voice app in 2019, to discourse markers when creating cohesion between turns of text-based chat with our chatbot templates—I basically spend all day, every day, using linguistics to form connections and drive product strategy.

What was the transition from university to work like for you? 

It was a huge culture shock! My first role outside academia was as a UX researcher at Microsoft. I was coming off the heels of having a limit on the amount of photocopies I could make in the Department of Linguistics to entering an office where food was catered every day, Post-Its could be used ad nauseam with nary a pang of guilt, and every time I had a research finding to articulate, it had to be done in a slide as opposed to a fleshed-out paper (to say nothing of my joy at being able to photocopy in color at the office). Everything required fewer words, had to be finished in a fraction of the time, and had to look way more visually appealing than anything I was ever expected to do in academia. I had to go beyond simply stating, “this is what I found in the data” and evolve towards, “we as a business should do y in order to make a market impact of z because the data say x.” (In industry, always lead with the “so what” and follow up with the data—something I learned from profoundly expert and gracious women in UX who taught me how to adjust to corporate priorities, read the atmosphere of a business, and succeed at advocating for oneself and one’s craft at the highest levels. I wouldn’t be where I am now without them.)

Do you have any advice do you wish someone had given to you about linguistics/careers/university?

My advice would be that your title and your degree aren’t the defining characteristics of who you are and what you do. When I was in academia, I deeply enmeshed my sense of self and self-worth in my status as a graduate student, and when I took a break after the MS, I had to mourn the death of a huge part of my identity. I had no idea who I was without the Department of Linguistics at Georgetown. But, after a lot of therapy and soul-searching, I came to the realization that even if I didn’t have the title of “student” anymore, that didn’t mean I was no longer capable of learning, just like even if I no longer had the title of “Director of Conversation Design,” it wouldn’t mean that I would suddenly become incapable of leading product strategy for conversational AI. Look inward. Titles, accolades, insignia—they’re distractions. Figure out what truly motivates you, what fulfills you and align to the elements of those qualities that present themselves in whatever role you pursue.

Any other thoughts or comments? 

A quick overview of the company I work for and the evolution of its conversational technology: Salesforce is a publicly traded company that provides customer relationship management software as a service in the cloud. In 2017, the company released the Einstein Bot Builder, a declarative platform for Salesforce administrators to create and deploy text-based chatbots to their own customers for service use cases. Since then, it’s grown into an expansive feature ecosystem, ranging from chat analytics to reusable chatbot templates. The Bot Builder, coupled with Slack, which Salesforce acquired in 2020, opened up the opportunity for me to establish a design practice by which Salesforce can create unified, consistent, and inclusive conversational chat experiences based on sociolinguistic research.

Recent interviews:

Interview with a Research Scientist

Interview with a Language Engineer

Interview with a Natural Language Annotation Lead

Interview with an Artist

Resources:

The full Linguist Jobs Interview List

The Linguist Jobs tag for the most recent interviews

The Linguistics Jobs slide deck (overview, resources and activities)

Conversation architecture is how you structure messages, roles, memory, and “trusted vs untrusted” text so the model follows your intent over many turns without drifting, slowing down, or getting hijacked. If you want reliable influence in AI‑mediated dialogue, you don’t win by writing longer prompts, you win by designing a repeatable conversation shape that preserves priority, preserves budget, and preserves control.

You’ll get a set of field-tested patterns for long-context prompting, system-prompt sizing, instruction placement, memory design, and prompt-injection resistance. The goal is operational: better adherence, faster runtimes, lower token burn, fewer regressions, simpler debugging, and fewer security surprises.

How Do You Structure An AI Conversation To Reliably Influence The Model Without Making It Brittle?

Start by treating the conversation like a product interface, not a one-off prompt. You need a stable “contract” that rarely changes, plus a per-turn payload that changes constantly. The contract belongs in the highest-priority channel you control, and it should stay short: role, non-negotiables, safety boundaries, output rules, and any definitions that must never drift. The payload belongs in a consistent per-turn template: goal, inputs, constraints, and the deliverable format.

Influence becomes brittle when instructions compete with each other or when the model can’t tell what is policy versus evidence versus user preference. Brittleness also shows up when every turn adds more rules, more exceptions, more “also do this,” until the instruction set contradicts itself. Keep the contract small, then implement “governance points” inside the turn structure: ask for a brief restatement of the task when the request is high risk, confirm assumptions when inputs are underspecified, and require structured outputs when downstream systems depend on stable formatting.

For multi-turn work, separate “what you want” from “what you know.” Put objectives and constraints in an instruction block, put supporting material in an evidence block, then request an output that explicitly references which evidence items were used. That single design choice reduces hallucinated glue, improves internal consistency, and makes review faster. It also upgrades debuggability, when a response fails, you can pinpoint whether the contract was unclear, the payload was missing a constraint, or the evidence was low quality.

Where Should Instructions Go In A Long Prompt: Before Or After The Reference Text?

Put durable instructions where they will not be buried, then reinforce them near the end when reference text is long. Long-context behavior often overweights what appears early and late, and underweights what sits in the middle. When you drop a massive transcript or a dozen documents into the middle, any directive placed there becomes easy to ignore in practice, even if it’s technically “in the window.”

A reliable layout for long prompts uses an instruction header and a directive footer. The header sets the task, the rules, definitions, and the output format. The body holds the reference material, clearly labeled as evidence or retrieved text. The footer repeats the deliverable requirements in a compact “do exactly this” recap, and it can include a short checklist that forces compliance: required sections, citation style, constraints, refusal rules, and any scoring rubric you want the model to optimize.

If the content is extremely long, stop thinking in terms of “one perfect prompt.” Chunk the evidence, run intermediate extraction turns, and carry forward only structured artifacts: key facts, contradictions, entities, open questions. This protects your token budget and improves attention on what matters. It also creates auditable intermediate steps that let you validate the model’s understanding before you ask for final synthesis.

Why Do Long System Prompts Reduce Accuracy, Speed, And Even Control Over Multi-Turn Chats?

Every token you keep in the system message consumes budget for user inputs and increases compute. On modern models, long prefill costs and key/value cache pressure show up as latency and cost. That’s the visible tax. The less visible tax is instruction dilution: the longer your system prompt becomes, the more you slip into redundant rules, overlapping constraints, and soft guidance that conflicts under pressure.

Control drops when the model can’t rank which instruction is decisive. A bloated system message also encourages teams to “patch” behavior by stacking more rules instead of fixing the underlying architecture. That’s how you end up with a system prompt that reads like a policy binder but performs like a coin flip. Tight system prompts outperform long ones because they stay internally consistent, remain readable to humans, and avoid stuffing the window with low-value text.

If you need deep behavior shaping, keep the contract short and move variable behavior into structured per-turn fields. If personalization is required, store it outside the prompt and inject only the minimal slice needed for the current turn. When the product demands strong consistency across sessions, design explicit state objects rather than repeating prose. Even research trends point toward compressing long prompt behavior into smaller control signals rather than hauling natural language rules through every request.

How Do You Prevent Prompt Injection And Instruction Hijacking In AI-Mediated Dialogue?

Start with a hard separation between trusted instructions and untrusted text. Retrieved documents, web snippets, user-provided files, and tool outputs must be treated as data, not commands. That separation must be visible in the message structure, the delimiters, and the model’s instruction hierarchy. If untrusted text can “sound like a system message,” the system is already compromised at the design level.

Operationally, implement three layers: labeling, filtering, and enforcement. Labeling means every evidence block is tagged as untrusted and never allowed to override policy. Filtering means running a sanitizer step that removes or flags instruction-like strings inside retrieved text, including “ignore previous instructions,” “reveal system prompt,” “exfiltrate secrets,” or tool-invocation bait. Enforcement means you give the model explicit conflict rules and require it to report detected conflicts, then proceed under the contract.

Defense pipelines also scale better than single-pass prompting. A multi-agent pattern is common in higher-security deployments: one agent classifies and strips hostile instructions, another agent extracts facts only, then the primary agent answers using the cleaned fact set. That reduces attack success while keeping utility. The important point is architectural: stop asking the same model instance to be judge, jury, and generator over raw hostile text, then acting surprised when it fails.

How Should Chatbots Handle Memory, Window Limits, And Conversation Summaries Without Losing Critical Details?

Stop relying on transcript accumulation as “memory.” Transcripts are logs, not state. Once you accept that, the design becomes straightforward: maintain a short recent transcript for tone and continuity, maintain a structured state object for the durable items. That state object holds user preferences, decisions, definitions, constraints, and open tasks. It also records what must not change unless the user explicitly revises it.

Summaries fail when they become narrative paragraphs. Narrative summaries hide omissions, bury constraints, and make it hard to diff what changed. Summaries work when they are structured and typed: bullets for decisions, a table for requirements, a list for unresolved questions, and a compact glossary for definitions. When you refresh the model, you inject the state object first, then the recent turns, then the user’s current request.

You also need refresh rules. When the user changes direction, the state object must update explicitly rather than drift implicitly. When the model produces a deliverable, record the output ID, assumptions used, and any constraints that were decisive. This improves continuity and reduces “policy creep” where every past decision becomes accidental law.

Should Retrieved Documents Go In The System Message Or The User Message?

The right answer is consistency plus trust signaling, not a single channel rule. Many teams get better adherence when retrieved text is inserted through a controlled template that the model learns to treat as evidence. That often pushes teams toward higher-priority channels for the wrapper, with the retrieved text clearly marked as untrusted. The win is not the channel itself, it’s the predictability of the wrapper and the clarity about what can and cannot issue instructions.

If you place RAG results in the same place as user instructions, you increase ambiguity. The model has to infer whether a sentence is a requirement, a quote, or a malicious override attempt. When you wrap documents in a stable evidence container, you reduce that ambiguity and make injection attempts easier to detect. You also make the system easier to test: prompt-injection red teams can target the evidence container and you can measure whether the model ever treats evidence as authority.

A practical compromise works well: keep immutable policy and formatting rules in system or developer messages, keep the evidence block in a dedicated section that is always labeled untrusted, and keep the user’s request as the only instruction payload for that turn. Then require the model to cite which evidence IDs support which claims. That forces grounding behavior without giving documents authority.

What Operating Rules Keep Multi-Turn Dialogue Stable Under Long-Context Load?

Define the “shape” of every turn and enforce it. A stable template reduces drift, improves evaluation, and gives you clean hooks for automation. The minimum viable turn shape is: objective, inputs, constraints, output format. When tool use exists, add tool permissions, tool output handling rules, and a schema for tool results that prohibits raw tool text from becoming instruction.

Implement checkpoints that cost few tokens and save hours of debugging. Require a one-line “task restatement” when requests are complex, require a short list of assumptions, and require a conflict report when evidence contradicts instructions. Those checkpoints prevent silent failure modes that look like “the model ignored me.” Most of the time, the model did not ignore you, it faced competing signals and made a choice you failed to constrain.

Measure performance the way an engineering org measures any interface: adherence rate, formatting validity rate, latency, token cost, safety failure rate, and user correction rate. Then run controlled changes. When a team edits the system prompt weekly, adds more policies, changes wrappers, and tweaks RAG formatting without tests, regression is guaranteed. Influence is not a writing contest, it’s a release discipline.

How Do You Structure A Prompt For Reliable Influence In A Multi-Turn Chat?

Short contract, stable template  

Trusted instructions separate from untrusted evidence  

Header rules plus footer recap for long inputs  

Rolling structured state, not transcript-only memory  

Build A Conversation Architecture You Can Ship And Defend

Reliable influence comes from message priority discipline, compact contracts, predictable turn templates, and explicit separation of commands from evidence. You keep chats stable by storing durable state outside the transcript and injecting it in a structured form, then limiting the system message to rules that must never change. You harden the system by treating retrieved text as hostile by default, filtering it, and enforcing conflict rules that always favor the contract. You also protect performance by controlling token growth, since slow, expensive prompts push teams toward shortcuts that break quality. Put these pieces together and the dialogue stops being a fragile prompt and becomes an engineered interface.

Conversational AI, or conversational artificial intelligence, refers to the use of artificial intelligence (AI) to enable natural language interactions between humans and machines. The goal is to create systems that can understand, interpret, and respond to user inputs (language) in a way that simulates human conversation. Conversational AI is employed in various applications, including chatbots, virtual assistants, and voice-activated systems.

Key components and capabilities of conversational AI include:

Natural Language Processing (NLP): NLP is a branch of AI that focuses on the interaction between computers and humans using natural language. It involves tasks such as language understanding, sentiment analysis, and language generation.

Speech Recognition: Conversational AI often involves the ability to recognize and understand spoken language. Speech recognition technology converts spoken words into text, enabling systems to process and respond to voice commands.

Intent Recognition: Systems need to identify the user's intent based on their input. This involves understanding the meaning behind the words and discerning the user's request or question.

Context Awareness: Conversational AI systems aim to maintain context throughout a conversation. This means understanding references, pronouns, and the history of the interaction to provide more coherent and relevant responses.

Dialog Management: Effective conversational AI involves managing the flow of a conversation, ensuring logical progression and addressing user queries in a coherent manner.

Personalization: Conversational systems can be designed to understand and adapt to individual user preferences and behaviors, providing a more personalized experience.

Multi-Modal Interaction: Some conversational AI systems support interactions beyond text, incorporating images, videos, and other forms of media to enhance the user experience.

If you stumbled across my page you are probably wondering why I migrated my website to Tumblr. The last time I used Tumblr as my main form of social media was in high school (circa 2013). During this time I was the most creative I'd ever been in my life - photographing my life, reposting quotes that led me to reading books, and so forth. Maybe it's the nostalgia factor, maybe it's my Leo tendencies to want to stand out (I apologize for the reference to horoscopes, I promise I only use them for comedic purposes), or maybe it's because I'm tired of paying for a platform that no longer inspires me. Using Tumblr to showcase my projects or learning has its own set of advantages, and were heavily considered before deciding to migrate. Whether it's better than a traditional website depends on your specific goals and preferences. Here are some benefits of using Tumblr

Ease of Use:

Tumblr is known for its user-friendly interface. It's easy to set up and customize your blog without the need for advanced technical skills.

You can quickly create and publish content, making it a convenient platform for sharing updates on your projects or learning journey.

Community Interaction:

Tumblr is a social media platform, and it allows for easy interaction with a community of users. You can gain followers, receive feedback, and engage with other users who share similar interests. I hope to reach an audience that perhaps has no clue what conversation design is. Linkedin, Medium, and other tech-focused platforms can seem intimidating.

The reblogging feature enables your content to be shared across the platform, potentially reaching a broader audience.

Visual Emphasis:

Tumblr is particularly visual-centric. If your projects involve a lot of visual elements such as images, graphics, or multimedia, Tumblr can be an excellent platform for showcasing them effectively.

The platform supports various media types, including photos, videos, and GIFs. This makes my job a little easier in that I don't have to rely solely on free image websites like Unsplash (sorry, ILY Unsplash).

Tagging and Discovery:

Tumblr uses a tagging system that helps organize and categorize content. This can improve the discoverability of your projects, especially if users search for relevant tags.

Tags make it easier for users interested in specific topics to find and follow your content.

Mobile-Friendly:

Tumblr is optimized for mobile use, making it accessible to users on various devices. This can be advantageous if you want your projects to be easily viewable on smartphones and tablets.

Cost-Effective:

Tumblr is a free platform, which can be beneficial if you're looking for a cost-effective way to showcase your work. You don't need to invest in hosting or domain fees.

Experimentation and Informality:

Tumblr has a more casual and informal atmosphere compared to traditional websites. This can be advantageous if you want to experiment with different styles of content or convey a more personal touch to your audience.

Screen shot taken from https://designguidelines.withgoogle.com/conversation/conversation-design-process/create-a-persona.html#create-a-persona-how-do-i-create-one

Googles ‘conversation design’ guide for third party developers wanting to make smart assistants. Google acknowledges that:

‘Users will project a persona onto your Action whether you plan for one or not. So it's in your best interest to purposefully design the experience you want users to perceive, instead of leaving it up to chance.‘