The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and analyze large volumes of untrusted data from a user-provided "corpus" (such as blog posts, emails, and chat history) to extract writing patterns.
Ingestion Points: The skill reads files from user-specified directories (e.g., ~/Projects/personal/javi.io/src/content/blog-es/) or accepts pasted text in chat interfaces.
Boundary Markers: The skill implements a robust three-pass extraction process (references/init.md) that includes a human-in-the-loop review (Pass 2 for alignment and Pass 3 for calibration). It also uses a GROUNDING.md file to explicitly define what facts can and cannot be asserted, preventing the agent from following instructions embedded in the data to hallucinate facts.
Capability Inventory: The skill utilizes file system tools including Read, Glob, Grep, Write, and Edit to build and maintain the voice profile.
Sanitization: The skill includes a dedicated "provenance triage" phase (references/corpus-qualification.md) designed to detect and exclude non-native content, such as AI-assisted text, translations, or clippings, which reduces the likelihood of processing adversarial instructions.
[DATA_EXFILTRATION]: The skill intentionally accesses sensitive user data, including personal emails, chat logs, and private writing, to build a style profile.
Evidence: The instructions in references/init.md and references/corpus-qualification.md encourage harvesting material from "sent mail, chat history, issue comments, and review threads."
Mitigation: The skill is designed to store this information in a local profile directory (.write-like-me/ or ~/.write-like-me/) and explicitly instructs the agent that "the profile never has to be published." There are no network exfiltration patterns detected; the data processing remains within the local agent environment.

write-like-me