git:20261105.c500fb5 to git:20261106.e3448e3

11 added, 5 removed. Audit A to A.

---

name: review-prompt

description: 'Review LLM prompts against the prompt-engineering skill''s principles — leading with where each line came from — and report the findings without modifying files. Use when reviewing prompt quality, auditing a prompt, evaluating a system prompt, or checking whether prompt issues are high-confidence and fixable.'

user-invocable: true

---

Review LLM prompts. Report findings; modify no file.

**First**: invoke the prompt-engineering skill to load the principles. Evaluate the prompt against them. They are opinionated defaults: where a rule the user stated, in the session or a context file, or a convention the project states or documents differs, that rule governs, a prompt following it raises no finding, and the report names the rule applied — except a High issue on the scale below that makes the prompt miss its goal, which is still reported, naming the rule that produced it.

- The provenance question leads: for each line, where did it come from — a user ruling, knowledge outside what the run will read, or a default it counteracts? A line the author could have worked out from material the run also gets is the finding. Judge a prompt by what its lines do, never by sections it lacks: a short prompt with nothing spare is the target, not a deficient one.

+ The provenance question leads: for each line, where did it come from — a user ruling, knowledge outside what the run will read, or a default it counteracts? Identify material the author merely derived from what the run also receives, and show the consequence of retaining it. Judge a prompt by what its lines do for its actual goal, inputs and consumer, never by sections it lacks: a short prompt with nothing spare is the target, not a deficient one.

+ The caller owns acceptance. When asked for a gate verdict, assess its substantive claims and any

+ deliberately required policy. A grade or removable sentence alone does not establish a material

+ failure; state the conflicting instruction, missing context or concrete behavioral or reading

+ cost. Required evidence that is missing leaves the assessment unresolved. Otherwise report the

+ findings and assessment without inventing a severity threshold.

**Input**: if no prompt is given (file path or inline text), ask before analyzing — don't assume.

Report format:

## Assessment: {Excellent | Good with Minor Issues | Needs Work}

**Strengths** — what's working, what the author should preserve.

**Issues**:

- | Issue | Severity | Fix |

- |-------|----------|-----|

- | {description} | High / Medium / Low | {concrete recommendation} |

+ | Issue and evidence of its consequence | Severity | Fix |

+ |---------------------------------------|----------|-----|

+ | {description and affected claim or concrete cost} | High / Medium / Low | {concrete recommendation} |

**Priority**: the highest-impact change first.

**Severity**:

- **High** — the prompt actively misbehaves or breaks a contract. Examples: contradiction between two rules that can't both hold; missing the goal entirely; absolute used on a judgment call that observably misfires; the agent declares a need for a tool it doesn't have, or omits a tool it actually uses.

- - **Medium** — the prompt works but drifts toward known failure modes. Examples: vague directive that produces inconsistent behavior across runs; restated model default adding noise the model has to wade through; a line whose only provenance is the author's own reading; a real gap left unclosed; text recording the prompt's own history — what a line used to say, when or why a rule changed, who asked for it; arbitrary numbers without a rubric; boundary failures — naming a harness-bound primitive, a rule-scope qualifier that silently excludes valid cases, mechanism stated as the only path, or one principle split across multiple places.

+ - **Medium** — a concrete quality gap materially impairs the prompt's use. Examples: an ambiguous directive produces inconsistent behavior; duplicated or historical instructions obscure which rule applies; an arbitrary number forces the wrong decision; a host-specific primitive excludes a supported context. Name that cost rather than grading a category of wording automatically.

- **Low** — polish or a nit of any kind: style, naming, structure, or otherwise. Example: duplication that adds reading load without changing behavior or the underlying promise.

Only flag high-confidence issues. Low-confidence findings are noise — skip style preferences, minor wording, and unverified hunches.