The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill downloads and executes an external CLI binary ('photocraft-cli') based on configuration found in 'references/commands.json'. The source is a third-party GitHub repository ('https://github.com/storytold/photocraft') that is not recognized as a trusted organization or well-known service, posing a high risk of supply chain attack.
[EXTERNAL_DOWNLOADS]: The 'bootstrap.py' script, executed during the initial setup in 'SKILL.md', fetches external dependencies and binaries from remote URLs that are not pre-verified or whitelisted.
[COMMAND_EXECUTION]: The skill relies on shell command execution via 'python3 -I -B' to run numerous internal scripts, including 'bootstrap.py', 'cli.py', 'workflow.py', and 'task_session.py'. These scripts are located in a harness directory ('$HARNESS_DIR') external to the skill's own controlled path.
[DYNAMIC_EXECUTION]: The skill implements a complex 'plan-based' execution model (e.g., 'craft-command-plan/v1'). It parses operations from JSON files and maps them to native CLI commands at runtime. This design permits the execution of any arbitrary command supported by the PhotoCraft CLI via manipulated JSON input.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits a significant attack surface for indirect prompt injection due to its processing of untrusted external data.
Ingestion points: The skill ingests complex JSON plans ('examples/poster-plan.json'), external native project files ('.pcraft'), and user-supplied assets via the '--asset' parameter.
Boundary markers: While the skill uses SHA-256 integrity checks ('expectedProjectSha256') for project state, it lacks robust natural language delimiters or 'ignore' instructions to prevent the agent from obeying malicious prompts embedded in image metadata or text layers.
Capability inventory: The skill possesses extensive capabilities, including filesystem modification, external CLI execution, and GUI interaction in 'bridge' mode.
Sanitization: The skill relies on structural validation of JSON keys and file hashes but does not sanitize the semantic content of instructions that might be embedded within the processed objects.

photocraft-cli-export