The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes session transcripts from the local filesystem, which are considered untrusted data and could contain malicious instructions designed to influence the agent's behavior.
Ingestion points: Transcripts are read during the triage and dimension-specific analysis phases as described in SKILL.md and prompts/analyst-common.md.
Boundary markers: Subagents are provided with a strict output format and explicit instructions to treat transcript content as evidence rather than instructions. Additionally, references/context-safety.md mandates bounded reading techniques to limit the amount of data processed at once.
Capability inventory: The skill can execute shell commands for analysis and packaging, write files to the workspace, and interact with the GitHub API.
Sanitization: A comprehensive redaction policy and scrubbing process are defined in prompts/scrub.md and references/redaction-policy.md to remove PII and secrets before export.
[COMMAND_EXECUTION]: The skill uses various shell utilities for data processing, file inspection, and bundle creation.
Evidence: Standard tools including awk, sed, grep, and jq are used for inspecting transcripts within references/context-safety.md. File archiving utilities zip and tar are used for bundle creation in SKILL.md. The GitHub CLI (gh) is used for managing issue reports.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to a well-known service to search for and create issue reports.
Evidence: references/github-issues.md details the use of gh and curl to interact with the GitHub API for searching and filing issues related to the project.

diagnosing-superpowers