The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands using user-supplied and agent-generated inputs without explicit sanitization.
Evidence: In Workflows/Harvest.md, the command bun ~/.claude/skills/_HARVEST/Tools/harvest.ts "<original input>" passes raw user input directly to the shell as an argument.
Evidence: In SKILL.md, the execution log command echo '{"ts":"'$(date...)","input":"8_WORD_SUMMARY",...}' >> ... incorporates an agent-generated summary into a shell command, which could be exploited if the summary is influenced by malicious content.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its processing of untrusted external content.
Ingestion points: Content is retrieved from arbitrary URLs via WebFetch and YouTube transcripts via fabric -y in Workflows/Harvest.md (Step 2).
Boundary markers: Absent. The skill does not define clear boundaries between instructions and untrusted data during the analysis phase (Step 3).
Capability inventory: The agent can append to local files (execution.jsonl) and execute a local script via bun (harvest.ts).
Sanitization: None. The workflow processes external data directly to extract ideas and patterns without filtering for embedded instructions.
[EXTERNAL_DOWNLOADS]: The skill inherently relies on fetching content from external network sources.
Evidence: The workflow utilizes WebFetch for articles and fabric for YouTube content as defined in Workflows/Harvest.md (Step 2).

Harvest