AI-assisted development (optional)

LLM mental model

A large language model is a parameterized neural network trained primarily to predict tokens from context. Weights/parameters are learned numeric values; they are not a database of retrievable sentences. A token is a model-specific text unit. The context window is the bounded input/output working sequence available to an inference. Generation repeatedly samples likely next tokens, so fluency is not evidence of truth.

Transformers use attention to compute context-sensitive relationships between token representations. Training changes weights; inference applies them. Fine-tuning changes model behavior through additional training. Retrieval supplies external material at inference time and does not itself update weights.

An embedding maps input to a vector where distance can approximate semantic relatedness. Vector retrieval finds candidates; it does not prove relevance, freshness or correctness.

Model and API vocabulary

• Family/tier: related models optimized for capability, latency or cost.
• Alias: a stable-looking name that may route to an updated model.
• Snapshot: a pinned version intended for more repeatable behavior.
• Reasoning effort: a runtime quality/latency/cost control on supported models, not a new set of learned weights.
• Temperature/sampling: controls token-selection distribution; zero can reduce variation but infrastructure and implementation details can still prevent perfect determinism.

Model names and availability age quickly. Verify them in official OpenAI model documentation rather than memorizing this repository’s snapshot.

Tools, MCP and skills

Tool/function calling lets a model emit structured arguments for code to execute; the application remains responsible for validation, authorization, execution and returning results.

MCP standardizes how an AI client discovers and invokes server-provided tools/resources/prompts. It is a protocol boundary, not an intelligence layer and not automatically safe. Treat server output as untrusted input and scope credentials/actions narrowly.

A skill is a reusable instruction/resource bundle teaching an agent a workflow. It may guide when and how tools are used; it is distinct from a tool, which performs an action. An agent combines model decisions, instructions, state and tools in a loop. Multi-agent systems add coordination cost and are useful only when work decomposes cleanly.

Productive coding workflow

1. State outcome, constraints, acceptance criteria and prohibited actions.
2. Give the smallest relevant repository context.
3. Ask for inspection and evidence before broad edits when risk is high.
4. Require tests/static analysis and inspect the diff yourself.
5. Evaluate repeated tasks on a representative suite, measuring correctness, review time, latency and cost.

Never paste secrets or unrestricted production data. Guard against prompt injection in issues, docs, web pages and tool results. AI-generated code inherits licensing, security, compatibility and maintainability review obligations.

Current OpenAI guidance distinguishes model aliases/tiers, reasoning controls, tool calling and multi-agent workflows; see official model guidance.

Interview prompts

• Contrast pre-training, fine-tuning, retrieval and prompt context.
• Threat-model an agent with shell, GitHub and database tools.
• Explain MCP to an engineer without calling it an API replacement.
• Design an eval for AI-assisted Laravel code review.