The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses sensitive local directories, specifically ~/.claude/projects and ~/.claude/usage-data/session-meta. These folders contain detailed session transcripts and user interaction history. This data is read by the DailyDigest.ts tool to summarize user activity for social media posts, which could potentially expose private development details if the summaries are published without review.
[COMMAND_EXECUTION]: The skill performs dynamic command execution using Bun.spawn in the DailyDigest.ts tool to run an external script (Inference.ts). Additionally, the GenerateVisual.md workflow executes a script from a separate directory (~/.claude/skills/Art/Tools/Generate.ts) to create images. The skill also allows for the execution path of the inference tool to be overridden via the XPOSTER_INFERENCE_PATH environment variable.
[PROMPT_INJECTION]: The UpdateAlgorithm.md workflow incorporates data fetched via WebSearch into the agent's knowledge base. This creates a surface for indirect prompt injection where malicious content in search results could influence the agent's tweet generation or scoring criteria. Mandatory evidence: 1. Ingestion: UpdateAlgorithm.md (WebSearch). 2. Boundary markers: None. 3. Capability: mcp__x-server__create_tweet. 4. Sanitization: Not mentioned.
[PROMPT_INJECTION]: The skill processes untrusted input from user-created files in the INPUT/ directory and automated session captures. Malicious instructions embedded in these source materials could potentially influence the agent's behavior during the content transformation and validation phases. Mandatory evidence: 1. Ingestion: BatchProcess.md (INPUT files and session captures). 2. Boundary markers: None. 3. Capability: mcp__x-server__create_tweet and Bun.spawn. 4. Sanitization: No explicit content filtering mentioned.

x-poster