The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves reading and analyzing 10-15 files from the user's codebase (Step 5) and then modifying them (Step 6). This ingestion of untrusted data from the local project provides a surface for indirect prompt injection, where malicious instructions embedded in the codebase could attempt to influence the agent's behavior during planning and implementation. The instructions lack explicit delimiters or warnings for the agent to ignore instructions found within project files.\n- [CREDENTIALS_UNSAFE]: An example project token (phc_jE9kXU0depRekiuabVROlxxkIXn95NqsNO3qB4qNKtl) is hardcoded in the BurritoConsiderationClient.xcscheme XML file within the references/EXAMPLE-swift.md reference. While PostHog project tokens (prefixed with phc_) are intended for public client-side use, their inclusion in reference materials is noted as hardcoded credential usage. Main skill instructions correctly mandate the use of environment variables for the user's actual project configuration.\n- [EXTERNAL_DOWNLOADS]: The skill installs PostHog SDKs from official package registries and references assets from posthog.com and its subdomains. These are legitimate resources owned and maintained by the vendor.

instrument-product-analytics