The Agent Skills Directory

[SAFE]: The skill implements a rigorous planning process with explicit 'gates' designed to prevent over-scoping and ensure technical accuracy. The 'Design Gate' specifically requires the agent to consider multiple alternatives and perform an adversarial-style review of its own plan.
[COMMAND_EXECUTION]: While the skill instructions discuss identifying verification commands (e.g., 'All existing tests pass with [command]'), it does not execute these commands itself. It is a planning-only skill that explicitly delegates execution to a separate process ('code-tasking').
[PROMPT_INJECTION]: The skill uses directive language ('The plan is incomplete unless...', 'Activate when...') to enforce structure and quality. These are functional constraints that improve the reliability of the agent's output rather than attempts to bypass safety filters.
[DATA_EXPOSURE]: The instructions budget for inspecting local project files, logs, and screenshots to provide context for the engineering plan. This is standard behavior for a development-focused AI skill and does not involve unauthorized access to sensitive system paths or credentials.
[INDIRECT_PROMPT_INJECTION_SURFACE]: The skill processes potentially untrusted data such as project logs, issue descriptions, and screenshots.
Ingestion points: SKILL.md (Evidence Budget section).
Boundary markers: The plan uses a strict structured Markdown format with defined headers to isolate different types of information.
Capability inventory: The skill's primary output is a Markdown document; it does not have autonomous file-write or network capabilities.
Sanitization: Multiple review gates (Scope Triage, Design Gate) require the agent to validate assumptions and classify boundary changes before including them in the final plan.

code-plan