The Agent Skills Directory

[COMMAND_EXECUTION]: The skill generates and instructs the user to execute a bash script (autodesign.sh) that runs the claude CLI with the --dangerously-skip-permissions flag. This flag is used to bypass all interactive confirmation prompts for tool use and file system access, enabling the agent to operate autonomously. While this is a functional requirement for the skill's purpose, it removes the human-in-the-loop safety mechanism for all actions taken by the sub-agents.
[COMMAND_EXECUTION]: The iteration script executes various external command-line tools, including playwright for web screenshots, codex for cross-model reasoning, and mobile development tools like maestro and detox. It also includes a potential deployment step using wrangler or other user-defined commands.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it ingests untrusted data from Figma (node names, design context, and visual content) and processes it to drive the design critique loop.
Ingestion points: Untrusted content enters the context through Figma node metadata and screenshots captured via Figma MCP or Playwright.
Boundary markers: The prompts utilize some structural separators but lack robust delimiters to isolate ingested design content from control instructions.
Capability inventory: The sub-agents have the ability to read and write files in the local directory, modify Figma files via MCP, and perform network requests through various capture tools.
Sanitization: The skill does not perform sanitization or validation of the data retrieved from external sources before including it in prompts.
[EXTERNAL_DOWNLOADS]: The skill defines dependencies on external repositories provided by the author and others, specifically jonmumm/skills@evals-first, jonmumm/skills@codex-review, and pbakaus/impeccable. These are fetched to support the evaluation and design refinement phases.
[DATA_EXFILTRATION]: The skill captures screenshots of the project UI and transmits them to external Large Language Model (LLM) providers (Claude and Codex) for visual evaluation. If the UI contains sensitive information, such as credentials or private data, it will be exposed to these external services during the critique process.
[REMOTE_CODE_EXECUTION]: The skill utilizes codex exec, which executes reasoning and logic through a remote Codex agent. Additionally, the shell script dynamically assembles and executes JavaScript code snippets within the Figma environment via the use_figma tool.

autodesign