The Agent Skills Directory

[DATA_EXFILTRATION]: The skill reconstructs session timelines by reading local agent logs and reporting findings to the 'Dyrits/SKILLS' public repository. This data transfer is the intended primary purpose of the skill. Security is maintained through explicit instructions to generically describe or omit private paths, code, client names, and secrets before drafting the report. Additionally, the skill mandates a human-in-the-loop approval step where the agent must show the exact body and wait for explicit approval before any issue is opened on GitHub.
[CREDENTIALS_UNSAFE]: The skill uses existing Git credentials stored in the environment helper for repository interaction. It specifically instructs the agent to use these credentials without printing them, which serves to prevent accidental exposure of tokens in the chat interface or logs.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from local agent logs, creating a potential surface for indirect injection if previous session data contains malicious instructions. The risk is minimized by the following evidence chain: 1. Ingestion points: Local agent logs (SKILL.md); 2. Boundary markers: The process requires establishing evidence-based findings through a step-by-step retrospective conversation with the user; 3. Capability inventory: Writing GitHub issues and creating local feedback files in '.agents/feedbacks/'; 4. Sanitization: Instructions require filtering out all private information and secrets, and applying 'unslop' guidelines to remove AI patterns.
[COMMAND_EXECUTION]: The skill utilizes Git commands and local file management to handle feedback records and report issues. These operations are within the scope of the skill's retrospective functionality and are performed under user oversight.

improve-skills