AI, Product & SDLC FAQ

Use AI to support one specific marketing task, such as drafting an offer, adapting approved content, or comparing campaign ideas. Keep your audience, evidence, and brand voice explicit. Judge the work by qualified conversations or conversions rather than the amount of content produced.

Start with a repetitive workflow that has clear inputs, a recognizable result, and an inexpensive way to check mistakes. Use AI for the part that requires interpreting information, while keeping exact rules in ordinary automation. Preserve human approval for actions with meaningful customer, financial, or operational consequences.

AI can help organize interview notes, propose themes, and draft follow-up questions, but it cannot replace contact with customers or turn generated opinions into evidence. Keep participant consent, source references, and contradictory observations visible. Treat a summary as a hypothesis you can check against the original material.

Evaluate an AI agent against the intended workflow, not just a polished demonstration. Check task completion, forbidden actions, recovery, latency, and cost on representative cases. Define which failures block release and who can approve a change in behavior before the agent reaches a wider audience.

Use AI to reduce the effort of testing a specific product assumption, not to build every feature in the imagined product. A useful MVP connects one audience, one problem, and one observable behavior. The prototype can be partly manual if that produces a clearer learning signal.

Put human intervention where a person can change a consequential decision before it becomes difficult to reverse. Useful checkpoints include uncertain intent, permission changes, external commitments, and recovery from repeated failure. Review should show the proposed action and evidence, rather than ask for a blind confirmation.

Use retrieval-augmented generation when the system needs relevant external knowledge at response time. Consider fine-tuning when you need to adapt model behavior through training examples. They address different problems and can be combined; first identify whether the failure is missing evidence, unclear instructions, or task behavior.

Measure AI ROI by comparing the value of a completed workflow with its full incremental cost. Define a baseline, a measurement period, and the quality threshold before launch. Include review, corrections, integration, and ongoing operation rather than treating lower token cost as the business result.

A useful specification prompt asks the assistant to expose uncertainty before it writes code. Include the user problem, intended behavior, constraints, and examples of success and failure. Ask for questions and testable acceptance criteria; do not ask it to invent business requirements to fill gaps.

Use AI as a practice partner that asks follow-up questions and critiques your reasoning against a consistent rubric. State your user, problem, alternatives, tradeoffs, and success metric aloud before requesting feedback. Practice quality is not the same as an employer’s interview score or a hiring prediction.

A dark factory is a metaphor for highly automated software delivery in which agents implement work from specifications and feedback. It is not a universally defined maturity standard. The useful question is which decisions can be delegated while requirements, evaluation, and operational accountability remain explicit.

A company brain is an informal name for a connected knowledge foundation that helps people and AI find relevant organizational context. It needs more than stored documents: ownership, permissions, source provenance, and updates determine whether an answer can be trusted for a particular decision.

LLMOps is the operational practice of developing, evaluating, deploying, and maintaining applications built around large language models. It becomes useful when prompts, model versions, retrieved knowledge, cost, or output quality can change the behavior of a workflow that people depend on.

An AI agent uses a model to choose actions and tools while pursuing a task. It is useful when the path cannot be fully prescribed in advance. A fixed workflow is often easier to operate when the steps, permissions, and success conditions are already known.

Context engineering is the work of selecting and organizing the information an AI system receives for a task. It includes instructions, retrieved evidence, tool descriptions, and relevant state. The aim is useful context within constraints, rather than the largest possible prompt or memory store.

An agent harness is the surrounding system that runs a model-driven workflow: it handles tools, state, execution limits, feedback, and recovery. The model is only one part of the agent. Harness design affects which tasks can be attempted and how success or failure becomes visible.

Agent evaluations test whether a model-driven system completes representative tasks within acceptable behavioral boundaries. They can examine the outcome, tool choices, permissions, and intermediate evidence. A useful evaluation set includes realistic success cases, failure cases, and situations where the correct action is to stop or ask.

Spec-driven development makes an explicit description of intended behavior a working input to implementation. With AI coding tools, the specification helps communicate scope, constraints, and acceptance conditions before code is produced. It remains a document that people must challenge, test, and revise when evidence changes.

Agentic workflows combine ordinary software automation with model-driven judgment for tasks such as issue triage, documentation, or change assessment. An agentic workflow should still have clear triggers, bounded permissions, auditable outputs, and a path for review when the result affects other people or production systems.

AI governance is the set of responsibilities, decisions, and controls that define how an AI system may be used and changed. A small team can start with use-case ownership, a data boundary, evaluation criteria, and an incident response path instead of creating a policy document nobody can apply.

An SRE agent can assist incident investigation by gathering operational evidence, proposing hypotheses, and suggesting remediation. Begin with read-only access and an accountable responder. Permission to inspect telemetry is different from permission to change production, restart services, or alter customer data.

LiteLLM provides interfaces and a proxy layer for working with language-model providers. A gateway can centralize aspects of request handling, authentication, routing, and usage controls. Whether it helps depends on your provider mix, operating needs, and willingness to maintain another component.

Model Context Protocol is a standard interface for connecting AI applications with external capabilities and context. It can make integrations more reusable, but it is not the same thing as a model, an agent, or a complete authorization system. Permissions and trust still need explicit design.

The Claude Agent SDK provides building blocks for running Claude-based agents with tools and context-management capabilities. It is useful when you need programmatic control over an agent workflow rather than only a chat interface. The application still owns its task design, permissions, evaluation, and operating environment.