ARC Skills
Engineeringintermediateupdated 2026-08-11v1.0

Prompt Ship Review

Review a system prompt against a ship bar for clarity, structure, ambiguity, injection surface, and named failure modes. Use when asked if a prompt is ready to ship, to review a system prompt, or to score injection surface and rewrites.

#prompts#injection#clarity#agents#security
View source in repo →
Install this skill

Copy the file into your agent runtime. Cursor loads personal skills from ~/.cursor/skills/prompt-ship-review/SKILL.md and project skills from .cursor/skills/prompt-ship-review/SKILL.md. Claude Code looks in .claude/skills/prompt-ship-review/SKILL.md.

Prompt Ship Review

When to use

A system prompt (optionally with a tool list) needs a structured review before production. Triggers: is this prompt ready, review this system prompt, check injection surface.

What it does

Scores the prompt against a fixed rubric and returns concrete rewrites. Aligns with ARC Labs Prompt Reviewer: clarity/structure, ambiguity phrases, OWASP LLM01-aligned injection surface, named failure modes, tool policy.

Steps

  1. Check for a clear contract: role, objective, must / never rules.
  2. Check structure: sections or delimiters so instructions and untrusted data do not blur.
  3. List weasel phrases ("be helpful", "when needed", "use your judgment") with a replacement for each.
  4. Check injection surface: instruction priority, trust boundary, untrusted-data framing, no secrets in-prompt.
  5. Name failure modes: refuse, clarify-or-stop, human gate for high-impact tools.
  6. If tools are present, require scoped descriptions and a gate on destructive actions.
  7. Emit letter-grade style findings plus paste-ready rewrite blocks.

Output

  • Overall: ship / ship-with-fixes / do-not-ship
  • Findings table: check, status (pass/warn/fail), evidence, rewrite
  • Failure modes: covered vs gap
  • One rewritten prompt skeleton if the original is salvageable

Guardrails

Scoring is heuristic. A high grade means the prompt *states* the right defenses, not that a live model is unbreakable. Never invent tool behavior that is not in the tool list. Do not put secrets, API keys, or production credentials into the rewritten prompt.