The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill performs automated analysis of project files, including README.md, package.json, pyproject.toml, and various source code files to understand the application context. This ingestion of untrusted project data constitutes a potential surface for indirect prompt injection.
Ingestion points: README.md, configuration files (package.json, pyproject.toml, requirements.txt), and project source code (.py, .ts, .js).
Boundary markers: The instructions do not define specific delimiters or instructions to ignore embedded commands within the scanned files.
Capability inventory: The skill gathers information to produce a text-based implementation plan; it does not execute code derived from the untrusted data.
Sanitization: No explicit sanitization or filtering logic is present for the ingested content.
[COMMAND_EXECUTION]: The skill instructions include a bash preamble designed to be executed by the agent to gather project context. This script uses shell commands like git and grep to inspect the environment.
[CREDENTIALS_UNSAFE]: The environment detection script accesses sensitive file paths to search for vendor-related configuration.
Accesses: .env and ~/.hindsight/config.
Purpose: The script checks for vendor-specific settings like HINDSIGHT_API_URL and HINDSIGHT_API_KEY to aid in the integration design.
[EXTERNAL_DOWNLOADS]: The skill provides implementation templates that recommend installing packages from official registries.
Packages: hindsight-client, hindsight-all, and @vectorize-io/hindsight-client.
These are identified as legitimate resources belonging to the skill's vendor, vectorize-io.

hindsight-architect