ce-doc-review

A document-review skill that critiques requirements or plan documents through multi-persona analysis. It dispatches generic sub-agents seeded with skill-local reviewer prompt assets, auto-applies safe_auto fixes, and routes the remaining findings through a four-option interaction — per-finding walk-through, auto-resolve with best judgment, append-to-open-questions, or report-only — for the user to decide. Use it when a requirements document or a plan document already exists and the user wants to improve it.

It runs in two interaction models. Default interactive mode presents the routing question, walk-through, and bulk-preview behaviors, and in Claude Code it pre-loads the AskUserQuestion tool before any question fires; a numbered-list fallback applies only when the harness genuinely lacks a blocking question tool. Headless mode, selected with mode:headless in the skill arguments, changes only the interaction model, not the classification boundaries: safe_auto fixes are still applied silently, while gated_auto, manual, and FYI findings are returned as structured text with their original classifications for the caller to handle, and the final phase returns immediately without routing or terminal questions.

A key step is classifying the document by its content shape rather than its file path. Requirements documents (what-to-build) are signaled by fields and headings such as actors, flows, acceptance examples, and identifiers like R1, A1, F1, and AE1, while plan documents (how-to-build) are signaled by implementation units, per-unit Goal/Files/Approach/Test scenarios fields, and U1-style identifiers. Path is only a tie-breaker, with ambiguous shapes defaulting to the more conservative requirements classification. Based on the document, conditional personas such as product-lens (challengeable premise claims or strategic weight) and design-lens (UI/UX, user flows, screens) are activated so each persona adapts its analysis to the document type passed through its template slot.