The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes repository data, product definitions, and project instructions which are potentially attacker-controlled, creating a surface for embedded instructions to influence agent behavior. * Ingestion points: The agent reads "repository instructions", "canonical product truth", and "current worktree" during the coordination phase (SKILL.md). * Boundary markers: The framework mandates a collaborative design phase with "explicit human acceptance" of the design package, and uses immutable digests to prevent silent tampering. * Capability inventory: The agent is permitted to invoke tools for research, spikes, and code execution (references/definition-checkpoint.md). * Sanitization: The lifecycle enforces role separation (e.g., the Implementer cannot act as the Verifier) and specialized owners judge evidence against a pre-accepted validation table.
[COMMAND_EXECUTION]: The instructions guide the agent to perform standard development operations such as installing packages, building software, and executing database migrations. * Evidence: The definition phase (references/definition-checkpoint.md) explicitly describes running commands to "install, build, generate, cache, migrate, or write reports" within isolated workspaces.
[DYNAMIC_EXECUTION]: The skill's delivery phase involves the generation and materialization of code patches and build artifacts that are executed as part of the verification and validation process. * Evidence: Isolated slice delivery (references/delivery-checkpoint.md) involves materializing binary patches or Git trees and running "technical verification" and "product validation" using tools like implementation-review and use-case-qa.

pursue-goal