The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill executes bundled TypeScript scripts, specifically references/consistency-check.ts and references/decide-access-scope.ts, using the bun runtime. The consistency checker performs byte-for-byte comparisons between different sections of the generated markdown files to ensure integrity, while the access scope script determines security levels for the modeled system.
[COMMAND_EXECUTION]: Several shell-based operations are integrated into the workflow:
curl is employed to download raw blueprint files from remote URLs for cloning and customization tasks.
yq is used to programmatically read from and write to the semantius/<org>/customizations.yaml file to persist user decisions across sessions.
bun is used to run validation scripts and the Semantius CLI tool (semantius call crud read_module) to inspect existing module states.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process complex, potentially untrusted data from remote markdown blueprints and user-provided requirements. This processed content directly influences the generated output and the logic flow of the agent. Although the skill includes an internal 'Audit' mode and consistency checks to mitigate malformed models, the ingestion of remote content remains a significant attack surface.
[PROMPT_INJECTION]: The instructions contain 'Narration restraint' directives and 'Interaction flow' logic that suppress internal state reporting. In certain modes (Guided/Fast), the skill is instructed to make decisions silently and 'never tell the user' or 'never mention them', which could be leveraged to hide automated actions from the user.

semantius-architect