The Agent Skills Directory

[PROMPT_INJECTION]: The skill generates and injects forceful, imperative directives into project files (e.g., STOP. You WILL forget skill names...) designed to override agent behavior and prevent context drift.
[DATA_EXPOSURE]: The inventory workflow (Step 1) involves reading directory listings from local paths such as ~/.claude/skills/ and ~/.claude/plugins/cache/*/agents/ to identify available tools.
[EXTERNAL_DOWNLOADS]: The skill references research findings and tool resources from Vercel's official blog and documentation.
[REMOTE_CODE_EXECUTION]: The skill optionally executes npx @next/codemod@canary, which downloads and runs code from Vercel's official package registry at runtime for Next.js projects.
[COMMAND_EXECUTION]: The skill uses shell commands including git log to analyze recent project history, ls for inventorying local directories, and npx for framework utility execution.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect injection by processing untrusted data from the local project environment.
Ingestion points: Project configuration files (e.g., package.json, go.mod, Cargo.toml), CLAUDE.md, and recent git commit history via git log.
Boundary markers: No explicit delimiters are specified to isolate untrusted data during subagent analysis.
Capability inventory: Shell command execution (git, npx), file writing (CLAUDE.md), and delegation to subagents.
Sanitization: No explicit sanitization or filtering of ingested project data is described.

reinforce-skills