The Agent Skills Directory

[PROMPT_INJECTION]: The skill's main instructions in SKILL.md employ strong instructional override language designed to alter the agent's default behavior. Phrases such as "Execute Discovered Obligations — do not acknowledge and stop" and "Do not stop with an apology, realization, praise of the guide, or a restatement... and do not hand the rule back to the user for confirmation" are used to enforce strict adherence to a complex technical workflow. While these are intended as operational optimizations for a development tool, they utilize the same techniques as malicious prompt injections to bypass conversational guardrails.
[INDIRECT_PROMPT_INJECTION]: The skill acts as an automated generator for game scripts based on user requests, which creates an inherent attack surface for indirect prompt injection.
Ingestion points: The agent accepts free-text requests for "Player attack" and "Player movement" skills as defined in SKILL.md and the Execution Core in references/execution-core.md.
Boundary markers: The skill relies on a rigid, stage-based execution framework to control agent behavior but lacks explicit delimiters or specific 'ignore' instructions for isolating raw user-supplied request text.
Capability inventory: The agent is empowered to create and modify project files, including components (.mlua), data structures (.userdataset), and documentation (.md), and to interact with platform tools like msw-search and msw-mcp.
Sanitization: The skill mitigates risks through a mandatory 'fail-closed' verification ledger (references/verification/) that requires concrete runtime evidence for every implemented feature, effectively gating completion on behavioral proof rather than just code generation.

maplestory-skill-maker