← MCP server and skill grades

skills

samber/developer-platform-skills/skills

A
Type
Skill
Scanned
Nov 13, 2026
Version
c042e5dc688e
Files read
315

This project is a set of guides, templates, and helper scripts that teach AI coding assistants how to design, review, and publish APIs, SDKs, and developer tools. The biggest risk is that one skill can tell the AI to edit your project's permanent instruction file without asking first, and a few sample workflow templates contain unsafe settings if copied directly into your release pipelines.

What it can reach

Files
Guides the AI and local scripts to read your project's docs, specs, and configs, and to create or update planning notes, changelogs, manifests, and the AI instruction file.
Shell commands
Instructs the AI to check git history, run spec linters or standard command-line checks, and run bundled Python scripts to verify versions and changelogs.
Network
Instructs the AI to read live vendor docs or test your public product links when web access is enabled, and includes optional Google Sheets scripts that contact GitHub if you manually copy and run them.
Browser
One guide instructs the AI to open your own public signup and login pages in an automated browser to check whether CAPTCHAs or bot blocks appear.

Findings

3 low
  1. LowKickoff skill can edit the project's persistent agent instructions without asking

    skills/developer-platform-kickoff/SKILL.md:287

    The kickoff router skill triggers broadly on API or platform tasks, reads repository git history, specs, and connected issue trackers, and allows the agent to patch the project's persistent agent-instruction file (such as CLAUDE.md or AGENTS.md) without asking for confirmation. Unlike the memory section in the same file, which requires showing a diff and getting approval first, this step has no confirmation gate and could persist unwanted or prompt-injected instructions into future sessions.

    Fix: Require showing a diff and obtaining explicit user approval before modifying any project agent-instruction file, and treat text from git logs and issue trackers as untrusted data.

  2. LowShell command injection risk via unquoted branch expression in CI workflow template

    skills/api-changelog-automation/references/tool-configuration.md:123

    The reference GitHub Actions workflow interpolates ${{ github.base_ref }} directly inside a bash run script instead of passing it through an environment variable. If a team copies this workflow into their repository and a pull request targets a branch whose name contains shell metacharacters such as $() or backticks (which Git allows in branch names), the runner will execute them as shell commands.

    Fix: Pass github.base_ref via an env: variable (for example BASE_REF: ${{ github.base_ref }}) and reference it in the shell command as "origin/$BASE_REF".

  3. LowUnpinned @latest scanner package executed with secret token in CI workflow template

    skills/agent-skill-offering/references/ci-release-gates.md:103

    The Gate 6 GitHub Actions snippet runs snyk-agent-scan@latest via uvx with SNYK_TOKEN exposed in the environment. Although a prose note below the snippet advises replacing @latest with a fixed version, an agent or developer copying the YAML block verbatim will download and run whatever package version is newest on PyPI on every CI run, exposing the repository and token to upstream supply-chain compromise.

    Fix: Pin a specific version in the YAML snippet itself instead of @latest, and pin actions/checkout to a full commit SHA.