The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands via the agent's environment to perform task management and file indexing. Evidence includes grep and ls operations:
grep -rn "^- \x5b \x5d" Projects/ Daily/ for finding open tasks.
grep -rn "due: 2026-04" Projects/ Daily/ for identifying overdue items.
grep -rl "keyword" Resources/ for searching across the vault.
ls Resources/{subtype}/ for directory listing.
[EXTERNAL_DOWNLOADS]: The skill integrates with a tool called WebFetch to retrieve content from external URLs provided by the user. While it explicitly forbids the use of other scraping services, fetching arbitrary web content introduces external data into the agent's context.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its highly autonomous data routing and rule-updating mechanisms.
Ingestion points: External data enters the context via WebFetch (Web Content Extraction section) and meeting transcripts/call logs (Meeting Intelligence section).
Boundary markers: The instructions do not define specific delimiters (like XML tags or clear markers) to separate ingested external content from internal instructions, nor do they instruct the agent to ignore instructions embedded within that data.
Capability inventory: The skill has extensive write capabilities, including appending to the 'Rules' section in claude.md, updating Context/voice.md, and modifying Context/business.md. It also creates and modifies project READMEs and daily notes.
Sanitization: There is no mention of sanitizing, escaping, or validating the content retrieved from web pages or transcripts before it is processed or archived into the vault. The 'Teaching Loop' specifically automates the addition of rules to claude.md without human intervention, which could be exploited if malicious instructions are interpreted as 'corrections' during the processing of untrusted data.

assistant