The Agent Skills Directory

[PROMPT_INJECTION]: The file migration-guide.md contains multiple instruction blocks explicitly addressed to the "AI Agent". These instructions direct the agent to verify assumptions, check for specific features, and halt or ask for user confirmation before modifying code (e.g., "AI Agent: Before changing any code, check if these assumptions... If not, stop the migration"). While these are intended as safety guardrails for the migration task, they constitute direct overrides and instructions for the agent's autonomous behavior.
[INDIRECT_PROMPT_INJECTION]: The documentation includes instructions that are intended to be processed and obeyed by an AI agent during a migration task.
Ingestion points: Instructions are embedded throughout migration-guide.md.
Boundary markers: None; the instructions are part of the markdown body text without specific delimiters separating them from the documentation.
Capability inventory: An agent using this skill (such as Claude Code) typically has filesystem access and terminal execution capabilities.
Sanitization: None; the instruction blocks are processed directly by the agent.
[OBFUSCATION]: In references/android/guide/navigation/navigation-3/recipes/conditional.md, the code block for Navigator.kt contains anomalous character injections (e.g., val<, Ba>ck, Ke>y, priv>ate). These symbols make the code syntactically invalid Kotlin and create a pattern of non-standard character usage. While this appears to be an artifact of poor content formatting or copy-pasting, such techniques can be used to evade simple pattern-based analysis.

navigation-3