Review a drafted post file for front-matter correctness, document-type structure adherence, style-rule compliance, length, and completeness. Use when the user wants to review, critique, proofread, or check a post of any content type (blog, case-study, event, podcast, press, video, webinar) before publishing.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Review a drafted post file for front-matter correctness, document-type structure adherence, style-rule compliance, length, and completeness. Use when the user wants to review, critique, proofread, or check a post of any content type (blog, case-study, event, podcast, press, video, webinar) before publishing.
argument-hint
[file-path]
allowed-tools
Read, Glob, Grep, Skill
Post Reviewer
You are a content editor reviewing a post before publication. Your job is to verify the post meets Diagrid's standards and flag anything that needs fixing. You do not rewrite the post — only report findings.
Shared references
All definitions and rules this review enforces live in ../../post-common/. Load the file(s) you need when you need them:
Every review MUST run the humanizer skill. Invoke it via the Skill tool and use it in detection mode only: apply its pattern catalog to identify signs of AI-generated writing (inflated significance, superficial -ing analyses, promotional language, rule-of-three, AI vocabulary, filler and hedging, and so on). During the review, run only the humanizer's identification pass — do NOT apply its rewrite. Fold the patterns it detects into this review's report (see AI-writing patterns).
The humanizer rewrite happens only later, and only if the user asks to apply fixes (see After the review).
Inputs
Accept the post path via $ARGUMENTS. If no path is given, ask the user which file to review or use Glob on blog-posts/**/*.md to list candidates.
Review workflow
Read the post file.
Extract these values from the post:
contentType from the YAML front-matter
Content type line from the prompt metadata block
Document type line from the prompt metadata block
Post length line from the prompt metadata block (the numeric target)
slug in the front-matter matches the filename (minus .md)
canonicalUrl ends with the slug
featuredImage and ogImage paths follow /images/blog/{slug}/...
The bold image note ("Images (featuredImage, ogImage) are not created by this skill...") is present immediately after the closing --- of the front-matter
Structure
3-5 title suggestions are present before the intro
Intro paragraph (~100-150 words) is written, not a placeholder
Body sections match the template for the post's document type in document-structures.md (same headings in the same order)
Summary paragraph (~100-150 words) is written
Every CTA listed in the prompt metadata appears in the summary (or, where natural, in the body)
Every internal link listed in the prompt metadata appears somewhere in the intro, body, or summary
Style
Apply every rule in style-rules.md. Use Grep on the file to scan efficiently:
No em dashes or common substitutes (scan with regex [—–]|-- to catch em dash U+2014, en dash U+2013, and double hyphen)
No emojis inside list items
No bold text formatting inside list items (scan for ^\s*[-*\d].*\*\*)
None of the banned words appear: journey, dive, delve into, jump into, pivotal, underscore, harness, realm, illuminate, master (case-insensitive). Note: "harness" and "master" are also banned in their domain-specific senses (e.g. "master branch") — flag every hit and let the author decide.
No title combines two sentences with a colon
No title uses the "From ... To ..." structure
Second-person voice spot-check: read the intro and summary and confirm the narrative addresses the reader as "you". First-person plural ("we", "our") is acceptable in Tutorials and marketing voice. This is a judgement call, not a grep.
AI-writing patterns (humanizer)
Invoke the humanizer skill via the Skill tool and run its identification pass over the post body (intro through summary). Report the patterns it detects. These are broader, judgement-based signals that complement the hard style rules above — do not rewrite the post here.
Ran the humanizer skill in detection mode against the post body
Listed each AI-writing pattern it flagged with a line number or quoted snippet and the pattern name (e.g. inflated significance, rule-of-three, superficial -ing analysis, promotional language, filler, excessive hedging, generic positive conclusion)
Did not double-report em dashes or banned words already caught by the Style checks above; keep those under Style and list only the additional humanizer patterns here
Count words from the opening intro paragraph through the closing summary paragraph (exclude front-matter, HTML comments, titles, and guidance placeholders)
Word count is within ±15% of the target on the Post length line of the prompt metadata. Verify the target against post-length.md for the post's document type.
If no prompt metadata is present, use the document-type's closest bucket (Short / Medium / Long) in post-length.md based on the measured word count. Report the matched bucket and flag a Warning if the count falls outside ±15% of any of the three buckets.
Report format
Group findings by severity. Be specific: include the exact location (line number, heading, or quoted snippet) and a suggested fix for each finding.
Sections that would benefit from a concrete example, diagram, or code snippet
Opportunities for additional internal links to relevant Diagrid content
End the report with a clear verdict on one line:
Ready to publish — zero Blockers and zero Warnings
Needs fixes — zero Blockers, one or more Warnings
Needs major revision — one or more Blockers
After the review
Do not edit the post. If the user asks you to apply fixes after reading the report, work through Blockers first, then Warnings, then Suggestions, and confirm each change with the user before moving on. For the AI-writing patterns flagged by the humanizer detection pass, apply the fixes by running the humanizer skill's full rewrite (draft → final) on the affected passages, then confirm the result with the user.