blog-review

Run the multi-agent quality review over a blog draft before it's marked ready: dispatch six parallel facet reviewers, confidence-score their findings, and consolidate one prioritized report. Use when

SKILL.md

GitHub

blog-review: multi-agent quality review

Review a blog draft against the rubric before it goes to reviewers. The shape is the same dispatch → confidence-score → consolidate as /code-review, adapted to blogs (blogs/CONVENTIONS.md §9), over six facets (a–f).

Two documents are the source of truth:

  • blogs/QUALITY.md — the criteria: the six facets and what each is scored against. This is what the reviewers grade against.
  • blogs/CONVENTIONS.md — the process this skill implements, plus the rules a draft must satisfy.

The review is advisory: it produces per-facet 0–100 scores and a prioritized findings report. It does not gate the frontmatter: the author resolves the findings and asks for approval in the review app. A human approval there is what clears status: ready, and merging that to main is the release.

Steps

1. Load context

  1. Read blogs/QUALITY.md (the rubric) and this skill's references/facets.md.

  2. Read the target index.md and its front matter: post_type sets the fact-density bar for facet (a), and audience is the yardstick for (c). Then read its brief, the yardstick for facets c and d (the draft is judged against what it set out to do). The brief is the first of these that exists:

    • blogs/<slug>/brief.md;
    • the ## Plan section of the PR that drafted the post: gh pr list --head blog/<slug> --state all --json url,body --limit 1.

    If none exists, say so in the report and judge (c) and (d) against the post's own thesis, without inventing a brief.

  3. Read blogs/tags.yml and blogs/SERIES.md — facet (e) checks against them.

2. Dispatch six parallel facet reviewers

Dispatch six reviewers in parallel (one message, multiple Agent/Task calls with the Explore or general-purpose type), one concern each, so no single pass has to hold everything. The six facets, defined in references/facets.md and scored against blogs/QUALITY.md:

  • (a) Sources & facts
  • (b) Voice & humanizer
  • (c) Audience & tone fit
  • (d) Structure & argument
  • (e) Links, metadata & series continuity
  • (f) Citability (human + AI)

Give each reviewer only its facet's slice of the rubric plus the draft, brief, and (for e) tags.yml/SERIES.md. Each reviewer returns a list (not prose) of flagged issues, each with: the location, the reason, and which QUALITY.md / convention rule it violates.

3. Confidence-scoring pass

For every flagged issue, score 0–100 for whether it's real and load-bearing (as in /code-review). Filter out false positives and nitpicks.

Apply the governing rule from blogs/QUALITY.md: voice (b) wins ties against citability (f). A facet-(f) finding that would flatten the author's stance is downgraded — say so explicitly rather than dropping it silently, so the author can judge.

4. Consolidate

Produce one report against index.md:

  • A per-facet 0–100 score (a–f) and an overall score, with a one-line rationale each.
  • Findings grouped by facet, ordered by confidence/severity, each pointing at a specific location with the rule it violates and a concrete fix.
  • State plainly that this is advisory: the author resolves the findings, then a reviewer's approval in the review app clears status: ready. Do not set the status yourself.

5. Publish-target check (conditional)

Only if the user names an HTML publishing target: run the Microdata / Rich-Results checklist from the "Publish-target concerns" section of blogs/QUALITY.md against the rendered HTML, not index.md. Skip this entirely for a draft-stage review — a Markdown draft can't carry Microdata and must not be scored on it.