The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted novel text provided by users, which represents a potential attack surface for instructions embedded within story data.
Ingestion points: User-provided text inputs and novel files processed via the /story-deslop trigger and scripts/deslop_guard.py.
Boundary markers: The system uses a "protection ledger" (references/fiction-protection-ledger.md) and explicit instructional boundaries to delimit story facts and character voices from editing actions.
Capability inventory: The agent performs local file system writes to apply candidates and executes diagnostic scripts.
Sanitization: The skill implements a multi-pass audit process, including a "Fidelity Audit" (Pass F), to verify that the agent does not alter plot points, clues, or character knowledge.
[COMMAND_EXECUTION]: The skill invokes several bundled Node.js and Python scripts to perform linguistic analysis, statistical metrics, and pattern matching on novel text. These are local, deterministic operations essential to the skill's primary function. (Evidence: scripts/language_gate.js, scripts/prose_metrics.py, scripts/check-style-hygiene.js).
[DYNAMIC_EXECUTION]: The skill includes logic to discover and spawn sub-agents (such as narrative-writer) based on configuration files found in project-specific directories like .claude/agents/ or .codex/agents/. This is used for platform-specific task orchestration. (Evidence: SKILL.md Agent compatibility section).

story-deslop