← MCP server and skill grades

skills

mbfinotti/advertising-skills/skills

B
Type
Skill
Scanned
Nov 13, 2026
Version
ab1eb51a2499
Files read
137

This collection of 31 markdown skills guides an AI agent through planning, auditing, and diagnosing paid advertising campaigns without making live changes in ad platforms. The main risk is that the kickoff skill permits unprompted edits to your project's global agent-instruction file, and the radar skill tells the agent to install a recurring background web-check routine without asking first.

What it can reach

Files
Reads project docs, campaign briefs, and user-supplied CSV or spreadsheet exports (including customer lists for audience sizing), creates or updates advertising-context.md and optional memory files, and may patch the project's agent-instruction file.
Network
When web browsing is available in the host agent, fetches ad platform help pages, public ad transparency libraries, public product reviews, user-supplied landing pages, and third-party marketing blogs, podcast feeds, and forums.
Browser
When a rendering browser is available, opens and inspects public landing pages at mobile and desktop viewports in a logged-off, read-only manner.
Shell commands
May read git commit history, propose git commits after showing a diff, or run local calculation and text-tokenizing scripts over exported reports when supported by the host agent.
Other
Reads and writes the host agent's persistent memory, detects connected CRM, analytics, and ad-account integrations, and creates or manages scheduled routines if the host agent supports them.

Findings

1 medium1 low
  1. MediumKickoff skill allows patching the project's global agent-instruction file without user confirmation

    skills/advertising-kickoff/SKILL.md:175

    While the skill description states that it creates or updates advertising-context.md, Section 5 also tells the agent to optionally patch the project's root agent-instruction file (such as CLAUDE.md or AGENTS.md) without asking permission or showing a diff first. Because the skill first reads existing campaign briefs, documents, and connected CRM or analytics sources, a prompt injection hidden in an untrusted brief or record could trick the agent into writing persistent instructions that affect all future sessions in that workspace.

    Fix: Require the agent to show a diff and get explicit user approval before modifying any agent-instruction file, matching the approval gate already required for git commits in Section 6, or limit writes strictly to advertising-context.md.

  2. LowRadar skill instructs the agent to install a quarterly scheduled routine without asking permission

    skills/advertising-radar/SKILL.md:157

    When a user asks for a reading or podcast watch list, the skill instructs the agent to directly install a recurring quarterly refresh routine whenever the host harness supports scheduled routines, without showing a dry-run or asking for approval first (unlike advertising-kickoff, which requires both). When that routine runs in the background, it automatically fetches external third-party URLs—including blogs, forums, and domains that may have expired and been resold—without an instruction to treat fetched web content as untrusted data.

    Fix: Require a dry-run preview and explicit user approval before installing the quarterly scheduled routine, and instruct the agent to treat all fetched external web pages and feeds as untrusted data.