View larger
Image by Zai, my ChatGPT, 2025
📌
Context: The Reality of Our Workflows
Many organizations rely on large workflow systems that route tasks to field reps and specialists around the world. These users often (in a day, or an hour):
Work across multiple accounts and assignments
Jump between disparate systems
Operate in time-sensitive, high-volume environments
A typical hour may involve:
Three or more browser windows
Ten or more open tabs
Multiple client accounts or projects
Switching between official tooling, messaging, email, and AI chat interfaces
In this environment, multitasking isn’t a behavior—it’s a necessity.
View larger
Image by Zai, my ChatGPT, 2025
View larger Currently, prompt sharing is informal and undocumented. Every use is an experiment, often based on someone else’s guess.
📌
The Ask
Everyone is being asked to engage with AI and find ways to use it to be more efficient, creative, and productive in their daily work.
And the most common, most reasonable first request people make is:
“Can we just start a prompt library?”
It’s a fair question.
It reflects real momentum.
And it deserves a structured answer.
View larger
Image by Zai, my ChatGPT, 2025
📌
The Quick Fix Everyone’s Reaching For: A SharePoint Prompt List
It seems helpful on the surface. But the brief history of GenAI has taught us that this is risky because it:
Lacks testing
Has no visibility into usage or failure
Encourages prompt drift and mutation
Ignores context boundaries
Puts the burden entirely on the end user
Why this matters: Without structure, even well-meaning prompt reuse quickly spirals into inconsistency, confusion, or compliance issues. In fast-moving, multi-tab, high-pressure environments, assumptions compound. We lose control not just of quality, but of accountability. Our customers and regulators don’t just expect good answers—they require us to explain our reasoning and show our work. AI outputs must be as transparent and traceable as any human decision.
It invites copy/paste culture—which is efficient, but brittle and unsafe in high-switch environments.
View larger
Image by Zai, my ChatGPT, 2025
Current general purpose chatbots lack the safety protocols necessary to provide different users consistent responses, even with the same set of prompts.
View larger The request is simple and reasonable: a central place to share and reuse helpful prompts. But without structure, governance, and testing, it invites inconsistency and risk.
📌
The Nature General-Purpose Chatbots
Many organizations begin their AI journey by using public general-purpose chatbots. These tools are accessible and powerful, but they were never built to handle the unique demands of internal enterprise work.
That’s why the next step is the creation of an internal enterprise chatbot —one that knows how to operate within the context of our systems, responsibilities, and data boundaries.
YourEnterpriseChatBot (and similar tools) are designed around a general-purpose Large Language Model (LLM) . These systems:
Don’t “know” your intent unless you tell them
Don’t “see” your workspace or account context unless given explicitly
Don’t enforce guardrails unless tools are layered on top
A general-purpose LLM is a precision tool with no situational awareness. Like a drill, it performs the task you request — not the one you meant .
Modern AI chat platforms depend on:
Prompt engineering frameworks
Tool integrations (e.g., search, file analysis, APIs)
Guardrails and risk filters
Identity and access management
Usage telemetry and observability
Without this scaffolding, a chatbot is just a context-blind text generator.
View larger
Image by Zai, my ChatGPT, 2025
The mission of most general purpose chatbots that you find on the web is to keep the user engage and not to provide accurate answers.
📌
The Human Risk: Muscle Memory and Context Collapse
FLOW STATE is when someone shifts from deliberate action to instinctive motion. When someone works quickly, they stop thinking in deliberate steps and begin relying on muscle memory :
Tab-switching becomes reflexive
Copy/paste happens instinctively
Prompts and responses start flowing between systems without proper boundary checks
The more windows and tabs are open, the more likely a mistake becomes inevitable. Examples of what can go wrong:
A prompt meant for one client is accidentally used on another
A copied response is pasted into the wrong workflow tool
A model-specific prompt is reused in a different system, creating misleading or risky output
This isn’t about carelessness. It’s about cognitive overload—and it’s universal. Fortunately, the AI overlords have figured this out and invented guardrails. These are automated checks and constraints—built into or layered on top of chatbots—that help prevent mistakes, filter unsafe outputs, enforce boundaries, and create accountability. They don’t stop the flow—they steer it.
View larger
Image by Zai, my ChatGPT, 2025
Great productivity comes from and maintaining flow state. However, when your tools do not support your flow state, accidents happen.
View larger
📌
What True North Might Look Like
The long-term vision (the steps to get us here are in the next section) is to create a purpose-built, safe, and efficient system for enterprise AI prompting. When possible, we want to built AI into our tools.
We recognize this is one of several possible approaches. The current state of the art provides the tools and design patterns to support a solution like this. We believe it offers the right balance between scale, safety, and how our people actually operate. But we also acknowledge there may be other valid models worth considering, and we remain open to learning, adapting, and evolving as we go.
This is how we scale from experimentation to trust. From clever hacks to sustainable systems.
Characteristics of True North
Context-aware copilots that understand user roles, assignments, and client contextIntegrated prompts tied to workflows, not floating in isolationGovernance-aware guardrails for red/yellow/green use zonesPrompt versioning and usage tracking Role-based prompt access and delivery Feedback loops and continuous prompt evolution Prompt compilation tools that translate human goals into structured AI behaviorsWorkflow integration , where the system—not the user—gathers all relevant data and passes it to the AI at the moment a prompt is invoked. When a user asks, for example, to “summarize the last three inspections for Account A,” the chatbot should automatically include those inspections in the prompt. No more copy/pasting between systems.
View larger In our long-term vision, prompts are invisible to the user—embedded in tools, protected by in/out guardrails, and always operating with context. Guardrails aren’t limits—they’re protections.
📌
The Middle Ground: A Curated, Tested, Human-Centered Prompt Library
Let’s meet in the middle—structure, safety, and agility.
🧩 Prompt Entry Includes:
Author and QA Owner (must be different people)Use case + model compatibility Tested examples Risk tier (Green / Yellow / Red)Last verified date Parameters , not hardcoded data
🧑🔬 Prompt Stewards
Assigned per department or tool
Triage new prompts
Monitor feedback
Prioritize updates
🔄 Feedback Loop
Rate prompts
Suggest improvements
Flag for errors or risk
Encourage reuse with attribution
Regular testing, especially when new models are released.
View larger
📌
How This Leads to True North
This curated library isn’t the final destination. It’s the on-ramp to:
Role-based copilots
Context-aware prompt delivery
Policy-driven guardrails
Workflow integrations
It gives us a foundation to:
Monitor real adoption
Identify needs
Spot risk patterns
Train future copilots
We’re not building bureaucracy. We’re scaffolding the future of responsible AI usage.
View larger
Image by Zai, my ChatGPT, 2025
Our tactical step forward: a curated library where prompts are tested, owned, and documented. This builds safety and learning into the process while we develop the full platform.
