The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from support ticket text and query results. • Ingestion points: Support ticket text and event property fields (e.g., $pathname, distinct_id) are used as data inputs. • Boundary markers: SKILL.md explicitly instructs the agent to treat all ticket content and query results as data, never as instructions. • Capability inventory: The skill uses read-only PostHog internal tools like posthog:execute-sql and posthog:experiment-get. • Sanitization: Detailed instructions are provided in references/pulling-the-data.md for escaping single quotes and backslashes to prevent SQL injection during data retrieval.
[COMMAND_EXECUTION]: The skill provides scripts/srm_check.py, a statistical analysis tool that uses only standard Python library modules (hashlib, csv, math, argparse, dataclasses, json, sys). It is intended to be executed by the operator or agent to validate Sample Ratio Mismatch (SRM) issues.
[DATA_EXFILTRATION]: The skill facilitates access to customer experiment data and identifiers. Privacy is maintained through strict access-check guidelines requiring verification of the requester's project permissions via the PostHog API before data is accessed.

debugging-experiments