Applied AI At Work / AI Primitives And Instructions
AI Primitives and Instructions
Learn the basic building blocks of AI applications: models, prompts, context, memory, tools, instructions, and evaluation.
Learn the basic building blocks of AI applications: models, prompts, context, memory, tools, instructions, and evaluation.
Use the brief to sharpen a real ai upskill conversation: what is the decision, what evidence matters, and what should remain human-led?
Capture one design rule you would reuse when reviewing an AI workload, assistant, or operating model.
Executive note
The Core Idea
AI applications are easier to understand when you can name their building blocks. These building blocks are often called primitives because they are the reusable pieces that many AI systems combine.
The common primitives are:
Once these are visible, an AI solution stops feeling like magic and starts looking like an engineered workflow.
- Model: the capability that interprets input and produces output.
- Prompt: the task instruction sent to the model.
- Context: the information available while the model works.
- Memory: selected information carried across interactions.
- Tool: an external capability the system can call.
- Guardrail: a policy, validation, or boundary that shapes behavior.
- Evaluation: the way quality and risk are measured.
Section 2 of 5
Prompt, System and Custom Instructions
Instructions exist at different levels.
System instructions define the broad operating rules: role, boundaries, safety expectations, formatting standards, and what the assistant should do when uncertain.
Custom instructions adapt the experience for a user, team, role, product, or workflow. They might set tone, preferred structure, business vocabulary, or expected depth.
Task prompts describe the immediate job: summarize this document, classify this request, draft a response, compare these options, or call a tool.
Recommended practice:
- Put stable behavior in system instructions.
- Put user or team preferences in custom instructions.
- Put the current job in the task prompt.
- Keep sensitive policy in system-controlled configuration, not in user-editable text.
Section 3 of 5
Tools and Actions
Tools let an AI system do more than generate text. A tool might search a knowledge base, query a database, create a ticket, calculate exposure, check a permission, or call another service.
Tool use changes the risk profile. A bad answer is one problem. A bad action is a larger problem.
Before allowing tool use, define:
- Who the model is acting for.
- What the tool is allowed to do.
- Which data the tool may access.
- Whether the action needs approval.
- How the action is logged.
Section 4 of 5
Memory and Personalization
Memory can improve continuity, but it must be intentional. Useful memory captures durable preferences, role context, or progress state. Risky memory stores sensitive details, temporary facts, or information the user did not expect to persist.
Good memory design gives users clarity and control. They should understand what is remembered, why it is useful, and how it can be corrected.
Section 5 of 5
Remember This
Most AI applications are combinations of the same primitives. When reviewing a solution, ask what model is used, what instructions guide it, what context it sees, what tools it can call, what memory persists, what guardrails exist, and how quality is measured.