name

review

description

Review English- or French-language written content against a fixed editorial checklist (plain language, idiomaticity, spelling, style/terminology consistency, broken links/anchors) and produce an annotated report with corrections. Use this skill whenever the user asks to "review", "proofread", "edit", "check", or "QA" a piece of writing — a document, article, webpage copy, help article, or draft — even if they don't name all five criteria. Also trigger for requests to check a document for broken links, inconsistent terminology, or awkward phrasing. Do NOT use for code review or grammar-only quick fixes on a short pasted sentence. Review French content in French, English content in English (US spelling).

Review

Audit written content against five fixed criteria and deliver an annotated report listing the issues found and suggested fixes.

Efficiency principle: the mechanical parts of this checklist (extracting links, checking anchors, flagging long sentences, detecting spelling-variant mixing, and validating URLs) are done by the bundled scripts — NOT by reading token-by-token. Run the scripts first, then spend model reasoning only on judgment calls (idiomaticity, plain language, terminology consistency).

The five criteria

Always report on all five, in this order. If a criterion is clean, say so explicitly — don't skip it, and don't invent issues to seem thorough.

Plain language — needlessly complex sentences, jargon without explanation, nominalizations, actor-obscuring passive voice, sentences the script flags as

30 words that could be split. (Plain English / plain French per the doc's language.)
Idiomaticity — phrasing that is grammatically correct but unnatural to a native reader (calques, stiff/literal constructions, wrong prepositions, unnatural collocations). Especially important if the text reads translated.
Spelling — misspellings and typos; plus US/UK mixing (the script lists variant pairs found). English content must use US spelling.
Style & terminology consistency — inconsistent capitalization, product/ feature naming, voice ("we" vs "you"), number/date/UI-element formatting. Build a mental glossary; flag any term used two ways.
Links & anchors — from the script output: broken/unreachable external URLs (by HTTP status) and internal anchors whose target heading doesn't exist. Note login-gated URLs as "could not verify", not broken.

Process

Get the content. Read the input file (.md / .mdx / pasted text) in full.
Run the deterministic pass from the doc's directory:
```
python3 scripts/extract_refs.py <DOC>            # compact report: anchors, long sentences, US/UK
python3 scripts/extract_refs.py <DOC> --urls | bash scripts/check_links.sh
```
Use --urls output (deduplicated) for link checking so no page body is ever downloaded. Only fall back to WebFetch on a specific URL when you must verify a redirect lands on a relevant page — that is the expensive path, use it sparingly.
Read once for the judgment criteria (1, 2, 4). Keep the categories separate: idiomaticity = naturalness to a native speaker; plain language = simplicity even for a native speaker. Tag every comment to exactly one criterion.
Produce the annotated report. Follow references/output-format.md and fill assets/report-template.html — do not hand-write CSS. Save to a _review/ folder at the root of the centreon-documentation repo.
End with the summary table: criterion, issue count, one-line takeaway (state "no issues found" per clean criterion).

Notes

Don't silently fix — the deliverable is a reviewable trail of what changed and why.
Tool name is WebFetch (Claude Code). If running in the claude.ai app instead, it is web_fetch and check_links.sh (curl) is unavailable — verify links with web_fetch there.

Review

Audit written content against five fixed criteria and deliver an annotated report listing the issues found and suggested fixes.

The five criteria

Always report on all five, in this order. If a criterion is clean, say so explicitly — don't skip it, and don't invent issues to seem thorough.

Plain language — needlessly complex sentences, jargon without explanation, nominalizations, actor-obscuring passive voice, sentences the script flags as

30 words that could be split. (Plain English / plain French per the doc's language.)

Idiomaticity — phrasing that is grammatically correct but unnatural to a native reader (calques, stiff/literal constructions, wrong prepositions, unnatural collocations). Especially important if the text reads translated.

Spelling — misspellings and typos; plus US/UK mixing (the script lists variant pairs found). English content must use US spelling.

Style & terminology consistency — inconsistent capitalization, product/ feature naming, voice ("we" vs "you"), number/date/UI-element formatting. Build a mental glossary; flag any term used two ways.

Links & anchors — from the script output: broken/unreachable external URLs (by HTTP status) and internal anchors whose target heading doesn't exist. Note login-gated URLs as "could not verify", not broken.

Process

Get the content. Read the input file (.md / .mdx / pasted text) in full.

Run the deterministic pass from the doc's directory:

python3 scripts/extract_refs.py <DOC>            # compact report: anchors, long sentences, US/UK
python3 scripts/extract_refs.py <DOC> --urls | bash scripts/check_links.sh

Use --urls output (deduplicated) for link checking so no page body is ever downloaded. Only fall back to WebFetch on a specific URL when you must verify a redirect lands on a relevant page — that is the expensive path, use it sparingly.

Read once for the judgment criteria (1, 2, 4). Keep the categories separate: idiomaticity = naturalness to a native speaker; plain language = simplicity even for a native speaker. Tag every comment to exactly one criterion.

Produce the annotated report. Follow references/output-format.md and fill assets/report-template.html — do not hand-write CSS. Save to a _review/ folder at the root of the centreon-documentation repo.

End with the summary table: criterion, issue count, one-line takeaway (state "no issues found" per clean criterion).

Notes

Don't silently fix — the deliverable is a reviewable trail of what changed and why.

Tool name is WebFetch (Claude Code). If running in the claude.ai app instead, it is web_fetch and check_links.sh (curl) is unavailable — verify links with web_fetch there.