The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process conversation history and extract learnings to be saved in durable memory. Because conversation history can include content from untrusted external sources (such as web pages or third-party tool outputs), there is an inherent risk of indirect prompt injection where malicious instructions could be saved as permanent agent rules.
Ingestion points: The skill instructions in SKILL.md (Step 2) require the agent to read back over the entire conversation history to identify decisions and corrections.
Boundary markers: The skill implements a 'safeguard' question ('Would I want this applied to every future job?') and a requirement to show a 'routing table' for user approval before writing, which serves as a human-in-the-loop boundary.
Capability inventory: The skill utilizes the agent's ability to read and write files to manage memory notes, project files, and instruction blocks.
Sanitization: While no technical string sanitization is specified, the skill uses a strict classification protocol (buckets) and a 'no-clobber' policy that surfaces conflicts to the user rather than overwriting existing data silently.

save-progress