Blog review

The half of shipping a post that the writer cannot do, because the writer has read every sentence six times and no longer sees it. This skill diagnoses and directs: it names the line, says what is wrong and what the fix is, and does not rewrite. A review that rewrites is a second draft with nobody checking it.

Two halves. The script catches what a grep catches better than a person. The read catches what only a person catches.

The script

.claude/skills/blog-review/scripts/check-post.sh <slug>

Run from the repo root. It exits with the number of blockers. What it checks, and why each one is there:

A warning is a judgement call the reviewer makes on the read. A blocker is not.

Some of the existing posts warn. That is information about the posts, not about the script: the rules were written after them, and a refresh is the moment to bring one into line.

The read

Read the post in the browser, not the MDX, at the width a reader gets. Then go through these, in this order, because the first ones are the ones the user rejects a draft over.

1. Does it read as written by a person? The shapes that give a draft away, all of which have been rejected by name ([[plain-marketing-copy]]):

• "Not X, but Y", "X rather than Y", "X instead of Y" as a flourish. The script counts these; the read decides which are a factual contrast (fine) and which define a thing by what it is not (cut).
• Uniform sentence length. Three sentences in a row of the same shape is the most reliable tell there is. The fix is usually one short sentence.
• A closing sentence that summarises the paragraph it closes.
• Superlatives and reassurance. "Exports that hold up" says nothing; "it stays sharp at 4K" says something.
• Wind-ups before a fact: "Research confirms the gap is worse than most expect." Cut the announcement and let the number do the work.
• Adverbs hiding a mechanism: "handles this seamlessly", "works transparently". Say what happens.
• Trade jargon where a plain word exists: "in post" for "by hand".

Do not send it to a humaniser. The firecrawl-blogs repo this skill is adapted from has one, and its remedy is colloquialisms, mid-sentence pivots and "in my experience". That is a different tell in a different voice, and this blog's voice is short, plain and directive. The fix for a machine-sounding paragraph here is fewer words, not more personality.

2. Is every claim about Prequel true? Check against the source, not the post before it:

• Spec figures against SPECS in LandingBody.tsx.
• What is automatic against ALREADY_DONE in the same file. Five things. Anything else described as happening on its own is an overclaim.
• Price and trial against lib/pricing.ts.
• Feature claims against FEATURE_ROWS in competitors.ts.
• Never cutting, trimming or silence removal. See [[no-dead-air-claim]].

3. Is every claim about somebody else's product checkable? A price carries a check date under the table. A Reddit, G2, HN or X citation links the post and names the subreddit or handle under the card; a card alone is not a citation, and a card whose words do not match the post at its link is a fabricated review attached to a named company. Open two of the links and compare the title word for word. A number with no primary source is cut. For the six in competitors.ts, the post agrees with that file, or the file is wrong and gets fixed first.

4. Is Prequel positioned as the skill says? Number one is the premise. No shortcoming named, no feature conceded. In a roundup, one factual Prequel win woven into every competitor section, true and checkable. In a guide, the other routes described fairly and briefly, then Prequel. Requirements stated once, near the pricing link, as a requirement.

5. Does the shape match the genre? [[guide-posts]] and [[comparison-posts]] each have one. The definitional question opens a guide and sits under the TL;DR on a roundup. The first paragraph under it stands alone. Sections lead on the action. Headings say what the section delivers, never "Overview", never two questions in a row.

6. Does the FAQ answer searches, not the article? Each question is one somebody types into a search box. Each answer stands alone, repeats the subject, names the price again if it names a price, and works Prequel in where it belongs. An answer that depends on the paragraph above it is an answer nobody sees.

7. Does the excerpt state what the post contains? Written as us, one field driving the description, the OG card and the standfirst. Not a pitch, not "Learn how to".

8. Read one section aloud. Any sentence that runs out of breath is two.

What is not checked here

Rankings, traffic and Search Console are [[seo]]. Site copy outside the blog is [[site-copy]]. A post's research is [[blog-research]]; if the review finds a price it cannot verify, that is a research task, not a review finding to guess at.

The report

## Review: <title>

### Blockers
- <file>:<line> <what is wrong>. Fix: <the change>.

### Warnings
- <file>:<line> <what is wrong>. Fix: <the change>.

### Passes
- <one line each, brief>

### Fix in this order
1. ...

Line numbers from the MDX, so the writer can jump. A clean section is said to be clean. Nothing is rewritten; the writer makes the change and runs the script again.