This skill auditsthe supporting relationship between conclusions and evidence, not rewriting business judgments on the user's behalf.pm-prd-review is responsible for document structure, feasibility, and acceptance criteria; this skill checks whether claims such as "market size, user pain points, competitor features, growth effects" are actually supported by sources. For formal determinations, follow the references/rubric.mdin the same directory; that file is installed with the single skill and does not depend on the full skill package.
Input and Boundaries
- Accept the report, excerpt, accessible document path, or material link provided by the user. When there is no material, ask only one most critical question: which material or which conclusions should be audited?
- By default,read-only and report in chat: do not write, modify, or overwrite source files by default. Only when the user explicitly asks to save, confirm the output path and overwrite risk.
- When the user saysno internet, use only the provided materials; do not access the internet, and do not write external sources as "verified". When materials are inaccessible, require login, or are unauthorized, mark them truthfully; do not fabricate sources, data, interviews, or screenshots.
- Do not upload the user's internal materials to third parties. Necessary retrieval is limited to public, non-sensitive claims; if the environment has no search tool, stop external verification and explain the limitations.
Audit Process
- Define scope: list the key claims that affect product decisions, numbered by their position in the original text
CLAIM-001、CLAIM-002...; for long documents, audit high-impact conclusions first, and do not pretend the full text has been covered. Record the original wording, metric definitions, time/region/population clearly. - Trace evidence: for each claim, find the original material cited, and verify the author/institution, publication date or data collection date, original values, and definitions. A source existingdoes not equalits supporting the current claim; paraphrased articles, summaries, and inferences need to be traced back to accessible original materials. Only further verify when the user permits internet access and tools are available; do not claim to have read the full text based on a search snippet.
- Handle contradictions and inferences: when credible sources conflict, present both sides and the differences in definitions side by side; do not pick a source that supports your own conclusion; causal, extrapolated, and predictive claims should have their premises separated out. Determine according to the rubric's fixed status constants, and give "why so" as well as what cannot be confirmed.
- Report next steps: for high-impact claims with weak evidence, give the minimal next verification action and suggested owner. The audit cannot automatically change
UNVERIFIEDintoVERIFIED, nor automatically modify the user's original text.
Output Format
First write the audit scope, review time, whether internet access was used, and uncovered parts. Then output a traceability table:
| Claim ID / original text location | Claim summary | Status | Source and date | Basis and limitations | Decision impact | Next verification action |
|---|---|---|---|---|---|---|
| CLAIM-001 / §2 | … | UNVERIFIED | Not provided; cannot be assumed | Unable to find data supporting this definition | High | Please provide the original report or adjust the wording |
状态Use only the enumeration constants defined by the rubric; do not substitute vague phrases such as "basically credible". Attach accessible links or file locations and dates for sources; for sources that do not exist, write "not provided" and do not fabricate them. Finally, listmay be retained, wording needs to be downgraded, evidence needs to be supplemented, conflicts existfour types of recommendations. For high-risk conclusions, explain which product decision needs to be re-reviewed if verification fails.
If the user requests direct revision of the source report, first separately confirm the scope of changes and retain the original text; the default output of this evidence audit remains a standalone report.