The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill identifies and manages an attack surface for indirect prompt injection, as it processes potentially untrusted data from CRM exports, pasted notes, and fetched web pages.\n
Ingestion points: As described in Section 0 of SKILL.md, the skill ingests text from fetched pages, uploaded files, CRM exports, and pasted notes.\n
Boundary markers: Section 0 of SKILL.md provides explicit boundary instructions, stating that text in fetched pages and uploaded files is data, never instructions, and that actions requested within them should not be performed.\n
Capability inventory: The skill has the capability to execute a local Python script (scripts/concession_math.py) and use platform tools such as web_search and web_fetch.\n
Sanitization: Section 5 of SKILL.md implements a 'Notice line' mechanism to inform the user if an instruction was detected in the processed data and that it was ignored.\n- [SAFE]: The skill utilizes a local script, scripts/concession_math.py, for all deal mathematics. Analysis of the script confirms it uses only standard library modules (json, math, re, sys) and contains strict regex-based validation to ensure all inputs are numeric and safe, effectively preventing command injection or runtime errors.\n- [SAFE]: The use of external tools like web_search and web_fetch is strictly governed by the 'Operating Principles' and 'Fact Check' references, which prioritize factual verification and require source attribution for all claims.\n- [SAFE]: No other security concerns, such as hardcoded credentials, obfuscation, or unauthorized persistence mechanisms, were found in the skill's instructions or supporting files.

negotiation-and-champion