The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests untrusted user feature ideas and reads local profile data to determine communication registers.
Ingestion points: User prompts (feature intent) and 02-DOCS/wiki/harness/user-profile.md.
Boundary markers: The skill enforces an "artifact is the contract" rule, requiring human approval of specifications and plans before any code is written.
Capability inventory: File writing to 02-DOCS/wiki/sdd/, subagent dispatching via the platform's Task tool, and Git operations (PR creation/merging).
Sanitization: Content is gated by ordered development phases; "Autopilot" mode, which reduces per-phase stops, is explicitly prohibited from performing destructive or outward-facing actions (merging, secret management, or deleting) without human intervention.
[DYNAMIC_EXECUTION]: The skill implements a model routing protocol that dynamically delegates development phases to specific model tiers (heavy, balanced, or light) resolved from a project configuration file.
Mechanism: Programmatic dispatching of subagents using platform-specific delegation tools (e.g., Claude Code's Task tool).
Constraints: This behavior is disabled by default (models.enabled: false) and requires local configuration in 02-DOCS/wiki/sdd/config.yaml. It is limited to resolving predefined model IDs for development tasks.
[COMMAND_EXECUTION]: The skill coordinates code implementation and shipping via sibling skills, enforcing project-wide authorship standards.
The ship phase enforces specific Git authorship ('Eric'), which aligns with the authorized vendor context ('ericrisco'), and prevents the inclusion of automated footers or co-authorship metadata that could misrepresent code provenance.

sdd