The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/sync-classify-prompt-from-core-product.mjs uses execSync to run shell commands, including go run and git commands, to update the skill's reference data. This script is intended to be executed in an environment where a specific external repository is present.
Evidence: Use of execSync with the go run ./tools/classify-prompt-export/ and git commands within scripts/sync-classify-prompt-from-core-product.mjs.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input in the form of human-provided text for classification, creating an indirect prompt injection risk.
Ingestion points: The content, assistant_message, and prev_message arguments in SKILL.md are used to provide the text for analysis.
Boundary markers: The snapshot prompt in PRODUCTION_CLASSIFY_PROMPT.md uses structured keys (e.g., content:"...") and explicit instructions to ignore auto-injected ambient blocks.
Capability inventory: The skill has access to the Bash tool and contains scripts capable of file system writes and command execution.
Sanitization: The skill's workflow includes steps to normalize whitespace and strip markdown elements like table rules and bold markers from the input before it is processed by the AI.

cawplan-internal-qa-coding-humaninputs