skills
mbfinotti/partnerships-skills/skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- 5764ab1dae47
- Files read
- 203
This project is a collection of 30 Markdown-based skills that guide an AI agent through planning, auditing, and managing channel partnerships, strategic alliances, affiliate programs, creator campaigns, and customer referral programs. It contains no executable code, and its main risk is that the kickoff router skill authorizes the agent to edit the project's global agent-instruction file without asking first, which could persist unwanted rules across future sessions if the agent reads a prompt-injected partner document.
What it can reach
- Files
- Instructs the agent to read workspace documents (including partner agreements, briefs, CSV payout exports, and git history), create or update partnerships-context.md at the project root, write persistent memory entries, and optionally edit the project's agent-instruction file.
- Network
- Instructs the agent, when the host tool supports web access, to fetch public pages such as user-supplied post URLs, creator profiles, competitor program terms, marketplace fee schedules, and regulatory guidance.
- Shell commands
- Ships no executable code or scripts of its own, but asks the agent to read git log history and, when code execution is available in the host, recompute commission payout spreadsheets or statistical baselines.
- Other
- Instructs the agent, when supported by the host harness, to inspect connected data sources (such as CRM or analytics connectors), persist non-sensitive project context in agent memory, and propose scheduled routines after user-approved dry runs.
Findings
-
LowRouter skill permits unapproved edits to global agent instruction files
skills/partnerships-kickoff/SKILL.md:189
The kickoff router skill triggers broadly on partnerships questions, instructs the agent to inventory workspace documents including third-party partner agreements and briefs, and authorizes patching the project's global agent-instruction file (such as CLAUDE.md or AGENTS.md) so every future session inherits those rules. Unlike the memory and scheduled-routine sections, this step does not require showing a diff or getting user approval first, nor is it mentioned in the skill description. If an inventoried external partner document contains prompt injection disguised as project constraints, the agent could write attacker-chosen instructions into the repository's global agent configuration.
Fix: Require showing a diff and getting explicit user approval before modifying any agent-instruction file, or keep all partnerships state confined to partnerships-context.md.
