Glossary

Agent — an LLM-based program the developer assigns tasks to, and which carries them out by producing and changing code.

Task setting — how the developer states to the agent what needs to be done and why.

Context — the data the agent sees while working: instructions, code, history, attached materials.

Specification — a description of what the system should do and why: scenarios, requirements, constraints, and acceptance criteria. The technical solution belongs in the plan, not the specification.

Plan — a description of how to implement the specification: the architectural approach, affected parts of the system, work sequence, and ways to verify it.

Context window — the limited amount of text and data a model can take into account in one session.

Skill — a saved procedure for an agent: instructions and, when needed, scripts, templates, and reference material for a recurring workflow.

Subagent — a separate agent instance with its own context, to which the main agent delegates a bounded part of the work.

Oracle — an independent source of a right-or-wrong answer: a test, linter, build, screenshot, reference result, or verifiable user scenario.

Testing seam — a system boundary through which behavior can be observed and verified without coupling the test to implementation details.

Tracer-bullet ticket — a small vertical slice of functionality that crosses the necessary system layers and ends in independently verifiable behavior.

Brownfield — an existing system with accumulated constraints, contracts, and decision history. Greenfield — a new system where those constraints do not yet exist.

SDD — Spec-Driven Development: an approach in which an agreed specification guides planning and implementation.

Pattern — a proven solution to a recurring problem of interacting with an agent.

Anti-pattern — a common way to make things worse: a tempting but harmful move that has a better replacement.