Glossary
Context Pack
A context pack is a bundle of product memory, artifacts, and constraints sent to an AI agent before it starts work.
Definition
A context pack is the agent-ready bundle that turns product memory into execution context. It can include CLAUDE.md, AGENTS.md, Cursor rules, UI guidelines, linked decisions, and the relevant ship plan.
Why it works
Prompting an agent from scratch repeats work and loses nuance. A context pack lets every run start with the latest product truth, so the agent understands intent before touching code.
What goes into a useful pack?
Begin with the task objective and the evidence required to call it complete. Add the relevant product decisions, customer insights, design guidance, technical constraints, and source links. The pack should explain both the intended outcome and the checks an agent must run before reporting a result.
Scope the material to the work. An onboarding change needs activation research and signup constraints; a billing change needs entitlement rules and payment behavior. Unrelated context consumes attention and can cause an agent to follow a rule that belongs to a different part of the product.
Example: handing off a search improvement
A PM can describe the search problem, link the customer evidence, identify permission boundaries, and specify what a good result should contain. The agent receives those materials alongside the implementation task. It can then explain how its changes meet the product intent and where uncertainty remains.
Before implementation, ask the agent to summarize the objective and identify missing information. After implementation, compare the result with the same acceptance criteria. This makes the handoff reviewable and helps distinguish completed work from a plausible-looking response.
How is a context pack different from a prompt?
A prompt expresses a request. A context pack provides the selected knowledge behind that request, with links that make it possible to check the source. Project-wide instructions can establish general rules, while a task-specific pack carries the decisions and evidence needed for one bounded piece of work.