Skip to content

Makes

Article write

Writes an article from a research brief, with links, metadata and schema, then checks it in 5 layers and stops for your signoff.

Other tools

Add this skill to Claude

Claude takes skills as a ZIP file. Your download has started; if it didn't, use the button below.

  1. 1In Claude, open Customize, then Skills.
  2. 2Choose +, then Upload a skill, and pick article-write.zip.
  3. 3Ask for it by name, or let Claude use it when a task matches.

Skills need code execution turned on in Claude's settings. On a Team or Enterprise plan, an admin can share it with everyone.

Download the folder

Try it now, without installing

See the skill work on a made-up input before you add it. Nothing is installed, and you can swap in your own input.

Open in the Claude app

Opens a new chat in the Claude desktop app with the prompt filled in. Nothing is sent until you press enter.

Includes the whole SKILL.md and the sample input. Paste it into Claude, ChatGPT or any chat.

Influenced by
Creating helpful, reliable, people-first content, Using generative AI content
Needs
The research brief · A named author with a bio page
Works with
Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md
Version
1.0.0 · Last checked · License MIT
Maintained by
Sabih AhmedContent
SkillContent

Article write

Example output

Checks before signoff
LayerResultOpen
Mechanical scriptPass1 warning
Claim checkPassNone
Fresh-eyes reviewerPass with notes1 for the author

Writes an article from a research brief, with links, metadata and schema, then checks it in 5 layers and stops for your signoff.

The skill

The whole skill, in plain text.

Nothing is hidden. Read it, change it to fit your team, or take the whole folder.

Scripts
Optional Python scripts for the counting. It works without them and says which way it ran.
Network
Opens web pages or your tools if your agent can. Works from pasted data if not.

3 test cases

article-write/SKILL.md

In the folder

  • article-write/SKILL.md
  • article-write/references/citations-and-links.md
  • article-write/references/meta-and-schema.md
  • article-write/references/review-rubric.md
  • article-write/references/writing-rules.md
  • article-write/scripts/article_checks.py
---name: article-writedescription: >-  Writes a B2B SaaS article from a finished research brief: outline, draft in the company's voice, citations,  internal links, title, meta description and JSON-LD, then checks its own work in layers (a QA pass, a script for  everything countable, voice-check, claim-check and one reviewer with fresh eyes), fixes what failed in at most 2  rounds, and stops for the reader's signoff with every open item labelled. Use it after article-research returns  READY or READY WITH GAPS. Not for choosing or researching a topic (use topic-precheck and article-research) or for  short posts (use founder-post or repurpose).license: MITcompatibility: >-  Works in Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md. Opens sources and the site's pages if the  agent can browse; otherwise works from the research brief and pasted pages. Python 3 is optional, for the  mechanical checks. A subagent or second chat is optional, for the fresh-eyes review.metadata:  author: SHAKEUP  version: "1.0.0"  stage: make  plugin: content  homepage: https://www.shakeuphq.com/skills/article-write--- # Article write Turns a research brief into a finished, checked article, so the reader's signoff is about judgment, not typos, broken links or unsourced figures. ## Purpose - One job: one article, with its metadata, schema and notes, ready for the reader's signoff.- Stage: make. It ends in its own layered check and a fix loop of at most 2 rounds.- It never publishes, never invents a figure, quote, customer, experience, author or credential, and never ships a sentence whose claim failed a check: it cuts the sentence and logs the gap. ## Use when / not | Situation | Use || --- | --- || `article-research` returned READY or READY WITH GAPS | this skill || Rewriting an article from a fresh brief | this skill || No research yet | `article-research` (and `topic-precheck` before it) || Checking a draft someone else wrote | `voice-check`, `claim-check`, `eeat-check` || Editing a live page from an audit | `refresh-audit` for the edit list, then this skill for the sections it touches || Posts, emails or scripts from the finished article | `repurpose` | ## Inputs | Input | Required | Ways to provide, in order | If missing || --- | --- | --- | --- || The research brief | yes | an `article-research` handoff and brief; a brief the reader pastes | `BLOCKED`: ask for it, or run `article-research` || A named author with a bio page | yes | the precheck or research handoff; `## Content` > `Authors` | `BLOCKED`: never invent a byline || The pillar and link targets | no | the precheck and `cluster-fit` handoffs; the sitemap if browsing; pages the reader lists | links from the brief only; flag "internal links not checked" || Voice rules | no | `## Voice` in the profile | write plain and direct; coverage `partial` || Approved claims | no | `## Claims and proof` | no first-party figures at all || Product mention rules | no | the reader's request; `## Product`, `## Positioning` | no product mentions | ## Profile read From `company-profile.md`: `## Company` (website, brand terms), `## Product`, `## Buyers`, `## Buying moments`, `## Positioning`, `## Messaging house`, `## Voice`, `## Claims and proof` (with `### Never say`), `## Competitors` (mention policy), `## Content` (authors, site sections), and every `## Learnings` entry for `article-write`, `voice-check` and `claim-check`. No profile: ask in one message for the website, the author's name and bio link, 3 to 5 words for the voice, and any figures the company can publish with their sources. Say that `company-profile` saves this next time. ## Steps 1. **Apply learnings.** Accepted rules override the defaults (for example "no definition box on our blog", "never mention pricing"). 2. **Gate the brief.** `COMMODITY`, `BLOCKED` or no brief means `BLOCKED`. Carry forward: primary keyword, angle and its evidence, verified sources, suggested sections, pillar, author, YMYL flag, and every gap. A gap stays a gap in the draft. 3. **Outline.** Apply `references/writing-rules.md` section 1: fix the parameters, then the H2 structure (active intent first), formats, key points, E-E-A-T and citation flags, FAQ, product mention points and link positions. The outline goes in the output for the reader; it is not a stop. 4. **Draft.** Apply section 2 in the profile's voice: title, introduction (tension in the first 2 sentences), key takeaways, the sections, definition box and pull quote where they earn a place, FAQ, close. Every figure from the brief's verified list or `## Claims and proof`. Experience lines only where the brief or the author backs them. 5. **Citations.** Apply `references/citations-and-links.md`: each external claim linked once at its sentence, publisher named, no links in headings or the introduction. A figure without a verified source is cut or reframed, and logged. 6. **Internal links.** Same file: live pages only, type caps, placement after the first 10%, anchors of 2 to 6 words that name the target's topic. The pillar link is required for supporting content. List up to 5 pages that should link here after publishing. 7. **Metadata and schema.** Apply `references/meta-and-schema.md`: title tag and meta description (a primary and 1 alternative each), slug, JSON-LD that matches the visible page, and the schema recommendations note. 8. **QA pass.** Score the draft against `references/review-rubric.md`, plus these people-first checks: it says something the top results don't (the angle is stated); every figure traces to a source opened in this run; a real author is named; the modified date changes only with real changes; nothing is written for AI crawlers instead of readers; the title, description, schema and alt text hold up as well as the body. Fix every critical. 9. **Mechanical checks.** Run:    ```   python3 scripts/article_checks.py --article article.md --profile company-profile.md \       --domain <site domain> --require-link <pillar URL> --require-schema   ```    Without Python, check by reading, in this order: dashes; words to avoid; leftover markers (`[[...]]`, comments, TODO, `[CITATION NEEDED]`, `[blank]`); H1 count; link targets and weak anchors; the required pillar link; internal links against 2 per 1,000 words; paragraphs over 75 words; title 30 to 60 and description 70 to 160 characters; takeaways 3 to 5; FAQ 5 to 7 with answers of 3 sentences or fewer; JSON-LD that parses, has `@type` and matches the visible FAQ. Say which way you ran it. 10. **Voice and claims.** Run `voice-check`, then `claim-check`, on the draft and the metadata. Without them installed, apply inline: the profile's words to avoid, point of view, sentences over 30 words, and the 3 copy tests; then every figure and comparison against `## Claims and proof` or its verified source, every `### Never say` line, and the competitor mention policy. 11. **One reviewer with fresh eyes.** Give a separate subagent (or a new chat) only the draft, the research brief and `references/review-rubric.md`. It quotes the exact text for every finding. Without a subagent, reread cold from the brief and say the review was not independent. Items only a person can judge are marked `needs the author`. 12. **Fix and recheck** (the Loop below), then stop for signoff. ## Check and verdict Severity comes from each layer: `blocker` (script), `critical` (QA, voice-check, claim-check, reviewer), `warning`, `suggestion`. | Verdict | Condition (first match wins) || --- | --- || `BLOCKED` | No research brief; the brief is `COMMODITY`; no named author; or the reader asks to invent a byline, credential, quote, customer or result || `SIGNOFF WITH OPEN ITEMS` | After the loop: 0 blockers and 0 critical in the draft, but warnings remain, or open items need input (a gap from research, a quote waiting for clearance, YMYL expert review, `needs the author` items) || `READY FOR SIGNOFF` | 0 blockers, 0 critical, no warnings left that have a clear fix, voice-check and claim-check PASS or PASS WITH NOTES, reviewer pass or pass with notes, no open items needing input | A draft never reaches signoff with a critical in it: when a fix would need an invented fact, cut the sentence and log the gap instead. Named stops: `BLOCKED` asks for the missing input. Signoff: every run ends by stopping for the reader to approve, with the article, the open items and the check results. Nothing in between stops. ## Loop 1. Collect every blocker, critical and warning from steps 8 to 11 into one list, each with its line and source layer.2. Fix in order: blockers, criticals, then warnings with a clear fix. A fix changes only what failed; never add a claim to fix a finding.3. Rerun the script, `voice-check` and `claim-check` on the changed draft. Rerun the reviewer only if round 1 changed a section's argument.4. At most 2 rounds. Then hand over: anything still open is labelled `open` with its layer, line and what would close it. When the script and the reviewer disagree about something countable (a paragraph's length), trust the script. ## Learn Append to `## Learnings` when: - the reader rejects a reviewer finding or reverses a fix at signoff (a `proposal`, with their reason: "we do use 'platform' as our category term");- the reader makes the same edit at signoff twice (a `proposal` to make it a rule, such as "no pull quote");- the article's results come back, such as impressions, rankings, demos or citations after 90 days (a `fact`, with its source). Example: `- 2026-10-24 | article-write | proposal | Put the comparison table in the first H2 for compare-intent articles | Why: the reader moved it up in 2 of the last 3 signoffs | Status: proposed` ## Output ```Verdict: <READY FOR SIGNOFF | SIGNOFF WITH OPEN ITEMS | BLOCKED> · Coverage: <complete | partial (what's missing)> · Rounds: <0 | 1 | 2> Outline: <parameters, then one line per H2: heading, job, format> <article.md: front matter (title, description, slug, author), the article body, the JSON-LD block, then SCHEMA RECOMMENDATIONS> Checks| Layer | Result | Open || QA pass | <pass | pass with notes | fail> | <n> || Mechanical (<script | by reading>) | <pass | fail> | <blockers, warnings> || voice-check | <verdict> | <n> || claim-check | <verdict> | <n> || Reviewer (<independent | not independent>) | <pass | pass with notes | fail> | <n> | Notes (open items): <each with its layer, the line and what would close it>Links to add from other pages: <page → concept>Learnings applied: <list, or none>``` Then the handoff block. Then stop for signoff. ## Hand off ```handoffskill: article-writeversion: 1.0.0verdict: SIGNOFF WITH OPEN ITEMScoverage: completefor: <article title>outputs:  title: <H1>  slug: <slug>  primary_keyword: <keyword>  author: <name>  pull_quote: <the pull quote line>  verified_sources: [<url>, <url>]  checks: {blockers: 0, critical: 0, warnings: 2}next: [eeat-check, repurpose]open: [customer quote on <subtopic> waiting for clearance]``` `repurpose` reads `pull_quote` and `verified_sources`; `eeat-check` reads the draft and `author`. ## Failure modes | Problem | Do this || --- | --- || A verified source can't be reopened at write time | Keep the claim only if the brief recorded its excerpt; otherwise cut it and log it || The vague rewording of an unverified claim would still imply it | Drop the sentence entirely || The reader wants word counts | Explain that length follows from covering the key points; give the current length as information only || The profile allows no product mentions but the reader asks for a call to action | One conversion link in the final 20%, no product claims || The topic is YMYL | Mark the draft for expert review at the top of the notes, source every factual claim, state limits plainly; the verdict is at best `SIGNOFF WITH OPEN ITEMS` until the review is done || No subagent and no second chat | Run the reviewer pass cold from the brief and say it was not independent | ## Influenced by - Google's guidance on helpful, reliable, people-first content and on using generative AI in content creation, which put accuracy, a real author and original value ahead of how the content was made: https://developers.google.com/search/docs/fundamentals/creating-helpful-content and https://developers.google.com/search/docs/fundamentals/using-gen-ai-content- SHAKEUP's content engine: every countable check done by a script before a reviewer reads the draft, one reviewer with fresh context who must quote the text, gaps logged instead of filled, and signoff always left to a person.

How it thinks

Inputs in, a verdict out, and a stop only where you decide.

What we built

In our own content engine, a script catches everything countable before a reviewer reads the draft, so people judge the argument, not the typos. The reviewer gets fresh context and must quote the text, and signoff always stays with a person.

It reads your company profile if you have one, or asks for only what it needs. Then it works through its steps, checks its own work and ends in a verdict.

The steps

  1. 01

    Outline from the brief

    The first section answers the buyer's main question in full. Follow-up questions come next, in the order a reader needs them, each with its format and the proof it needs.

  2. 02

    Draft in your voice

    The tension lands in the first 2 sentences. Every figure comes from the brief's checked sources or your approved claims. A gap in the research stays a gap.

  3. 03

    Cite and link

    Each outside claim links once, where it's made, with the publisher named. Internal links follow caps per page type and always include the pillar for supporting content.

  4. 04

    Write the metadata and schema

    A title tag and meta description in range, with one alternative each. The JSON-LD matches what's on the page, with the same author as the byline.

  5. 05

    Check in 5 layers

    A QA pass, a script for everything countable, a voice check, a claim check, then one reviewer with fresh eyes who must quote the text for every finding.

  6. 06

    Fix twice at most, then stop

    Blockers first, then critical findings, then clear warnings. A fix that would need an invented fact becomes a cut sentence and an open item.

What you need

  • Required

    The research brief

    An article research handoff and brief → A brief you paste

  • Required

    A named author with a bio page

    The precheck or research handoff → The Authors line of your company profile

  • Optional

    The pillar and pages to link

    The precheck and cluster fit handoffs → Your sitemap, if your agent can browse → Pages you list

  • Optional

    Your voice and approved claims

    The Voice and Claims and proof sections of your company profile

What it can return

  • READY FOR SIGNOFF

    Every layer passes and nothing needs your input.

  • SIGNOFF WITH OPEN ITEMS

    Nothing critical is left, but some items need you, such as a quote waiting for approval or an expert review.

  • BLOCKED

    There's no brief, the brief says the topic is commodity, there's no named author, or you ask it to invent a byline, quote or result.

Where it stops for you

  • Every run ends with your signoff on the article, the open items and the check results

Example output

What you get back.

A run for a made-up company. Your output depends on your inputs.

Here's a run for a made-up B2B SaaS company that sells attribution software. The brief is on attribution reports in Salesforce.

Verdict

SIGNOFF WITH OPEN ITEMS

Coverage: complete

Rounds: 1

Outline

reporting pipeline by channel in Salesforce, why rewriting stages breaks trust, and Campaign Influence or a new object. Then setup time, what to ask a vendor, and an FAQ with 5 questions.

LayerResultOpen
QA passPass with notes1 thin section
Mechanical scriptPass after round 1Description was 163 characters, cut to 148
Voice checkPass with notes1 sentence over 30 words, split
Claim checkPassNone
Reviewer, independentPass with notes1 item for the author
Open items

the author confirms the example in the Campaign Influence section, and 2 older guides should link to this article.

The handoff block the next skill reads
skill: article-write
version: 1.0.0
verdict: SIGNOFF WITH OPEN ITEMS
coverage: complete
for: Salesforce attribution that leaves your stages alone
outputs:
  slug: salesforce-marketing-attribution
  pull_quote: The board doesn't want a better model. It wants a number RevOps won't argue with.
  checks: {blockers: 0, critical: 0, warnings: 1}
next: [eeat-check, repurpose]
open: [author confirms the Campaign Influence example]

What good looks like

  • Every figure traces to the brief or your approved claims
  • The pillar link is there for supporting content
  • The schema's author and FAQ match what's on the page
  • A checks table with all 5 layers and what's still open
  • A clear stop for your signoff

Install

Use it in the tool you already have.

Pick your tool. Each tab has the shortest path we know of.

  1. 01
  2. 02
    In Claude, open Customize, then Skills.
  3. 03
    Choose +, then Upload a skill, and pick the ZIP. Code execution must be on in your settings.
  4. 04
    Ask for it by name, or let Claude use it when a task matches.

Questions

Before you add it.

Does it set a word count?

No. Length follows from covering each section's points and stopping. It reports the length as information only.

What does the script check?

Dashes, words you avoid, leftover placeholders, the H1, link targets and anchors, the pillar link, internal link density, paragraph length, title and description length, takeaways, FAQ size and the JSON-LD. Without Python, the skill checks the same list by reading.

Who is the reviewer?

A separate subagent or a new chat that sees only the draft, the brief and the rubric. If your tool can't start one, it rereads the draft cold and says the review wasn't independent.

Will it ever publish?

No. It stops for your signoff every time, and it never touches your CMS.