// Load when reviewing agent definitions or skill content. Adversarial review methodology — finding quality, severity, and structured reporting.
Load when writing or reviewing agent .md files. Design guidance for agent definitions and meridian-specific conventions.
Load when writing, reviewing, or improving any prompt artifact. Research-backed principles across four levels: prompt craft, skill extraction, agent design, and multi-agent coordination.
Load when writing or reviewing skills. Design guidance for SKILL.md files and bundled resources.
| name | prompt-review |
| type | reference |
| description | Load when reviewing agent definitions or skill content. Adversarial review methodology — finding quality, severity, and structured reporting. |
| model-invocable | false |
Find what's wrong, not confirm what's right.
The writer already believes their agent works. Your value comes from challenging that assumption — the principle violation they didn't notice, the ambiguity that will confuse the model, the scope creep that dilutes focus.
Each finding: what's wrong (reference the line), why it matters (the failure mode), and what to do about it. Prioritize things that require understanding principles and intent over surface issues.
Lead with blocking issues.
How does this agent fail? What happens with ambiguous input, missing context, conflicting skills? Which principles does it violate and what breaks?
Check for LLM writing patterns that weaken prompts: overcorrected guidance encoding "stop doing X" as absolute prohibition, contrastive definitions ("not X — it's Y") that only make sense in a conversation, prescriptive checklists where a principle would transfer better, labeled conclusions that restate what the examples already showed.
