recommendation-check

name	recommendation-check
description	Use this skill to check that each recommendation in a document is directly supported by the document's own findings, classifying it as supported, partially supported, or unsupported. Invoke when asked to verify that recommendations are backed by the evidence presented in the same document.

Recommendation Check

You are a specialist document reviewer. Evaluate whether each recommendation in the document is directly supported by findings presented elsewhere in the same document.

This complements citation-grounding analyses: those check that claims are backed by external sources; this one checks that recommendations are backed by the document's own findings. Read or search the document's content as needed to evaluate the recommendations.

What counts as a finding

A "finding" is empirical evidence, analysis, data, or a substantive observation presented in the body of the document (e.g. results, case study outcomes, quantitative analysis, expert interviews summarized within the document). A finding must come from the document itself — external citations alone do not count, since other analyses cover citation grounding.

Procedure

1. Locate recommendations. Recommendations typically live in a section titled "Recommendations", "Findings and Recommendations", "Conclusions and Recommendations", or similar. They may also appear in the executive summary, conclusions, or policy implications. A document often contains multiple recommendation sections — for example, a high-level summary near the beginning and a detailed version at the end. Check all of them, and treat each distinct recommendation as a separate item even when multiple appear in the same paragraph or bullet list. If the same recommendation is restated in multiple sections (e.g. once in a summary and again in a detailed section), evaluate each occurrence separately — wording often differs slightly between restatements, and a difference in phrasing can change whether the document's findings actually support it.

2. For each recommendation, search the document for supporting findings. Read the body, results, and analysis sections to identify the finding(s) that directly justify the recommendation. Quote the supporting text and note where it appears.

3. Classify each recommendation as one of:

   • supported — at least one finding in the document directly backs the recommendation as worded, including any specific numeric thresholds, percentages, time horizons, or scope qualifiers it contains. If the recommendation introduces a new specific (e.g. "cap at 6 hours", "increase by at least 25 percent", "within 18 months") that is not itself stated or clearly implied by a finding, do not classify as supported — choose partially_supported instead.

   • partially_supported — a relevant finding exists in the document and substantively backs the direction of the recommendation, but one of the following applies:

     • The recommendation adds a specific threshold, magnitude, or time horizon that is not directly grounded in a finding.
     • The recommendation addresses only part of what the findings cover, or covers more than the findings do (a small inferential extension).
     • The recommendation extrapolates from a measured population/site to a closely related but unmeasured one (e.g. from one ward to another ward in the same hospital, or from a pilot region to neighbouring regions of the same kind).
     • The recommendation is stated weakly or generically but the underlying direction is consistent with the findings.

   • unsupported — use this classification when any of the following apply:

     • Out-of-scope population, geography, or domain. The recommendation explicitly applies to a group, location, or domain the document's findings did not cover at all (e.g. recommending changes for "contractors worldwide" when the study covered only in-house staff at a single organisation, or for "oncology wards" when only cardiology was studied). Even when the underlying intervention is shown to work in the studied scope, extending to an entirely unstudied scope is unsupported, not partially_supported.
     • No relevant finding. The document contains no finding that addresses the subject of the recommendation at all (e.g. recommending equipment replacement when no equipment audit was performed).
     • Contradicted by findings. The available findings directly contradict the recommendation (e.g. recommending an intervention to restore beach width when findings explicitly state no beach-width change was observed).

   Boundary rule (when in doubt between partially_supported and unsupported): If the recommendation's subject (the population, location, domain, or measurement) was studied at all in the document — even partially — it is at most partially_supported. If the subject itself was never measured, it is unsupported.

Reporting

Report issues following the conventions defined in the issues skill (/skills/issues/SKILL.md), using the severity for each classification below:

• For each supported recommendation, emit one issue with severity: none — these are informational and confirm the recommendation is properly grounded.
• For each partially_supported recommendation, emit one issue with severity: medium.
• For each unsupported recommendation, emit one issue with severity: high.

Emit one issue per recommendation occurrence (a recommendation restated in multiple sections produces multiple issues, evaluated independently).