Skip to content

Gets data

Article research

Reads what ranks for a cleared topic, maps what every top page says and what none of them say, and proposes the angle only your company can back.

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-research.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
Optimizing for generative AI features in Search
Needs
The cleared topic · The top results for the keyword
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 research

Example output

What the top results cover
SubtopicCovered byClass
Campaign Influence setup4 of 5Baseline
Keeping stages intact1 mentionThin
Setup time on SalesforceNoneGap

Reads what ranks for a cleared topic, maps what every top page says and what none of them say, and proposes the angle only your company can back.

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-research/SKILL.md

In the folder

  • article-research/SKILL.md
  • article-research/references/research-method.md
  • article-research/scripts/coverage_matrix.py
---name: article-researchdescription: >-  Researches a B2B SaaS article topic that has passed topic-precheck: reads what the top results actually cover,  builds a coverage matrix of what they all say and what none of them say, pulls the company's own evidence, lists  the questions buyers ask in their own words, checks every figure against its source, and proposes the one angle  only this company can write. Returns READY, READY WITH GAPS, COMMODITY or BLOCKED. Use it after topic-precheck  says GO and before outlining or drafting. Not for choosing the keyword (use keyword-selection) or writing the  article (use article-write).license: MITcompatibility: >-  Works in Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md. Reads the live results and sources if  the agent can browse; otherwise works from pages and sources the reader pastes. Python 3 is optional, for the  coverage matrix and source ages.metadata:  author: SHAKEUP  version: "1.0.0"  stage: get  plugin: content  homepage: https://www.shakeuphq.com/skills/article-research--- # Article research Turns a cleared topic into a research brief a writer can use: what ranks already says, what it misses, and the evidence to say something it doesn't. ## Purpose - One job: one research brief for one topic, with an angle backed by evidence found in this run.- Stage: get. It reads data in and structures it; it ends with a readiness verdict for the writer.- It never fills a gap with invented evidence, a remembered statistic or a source it didn't open. A gap stays a gap, labelled. ## Use when / not | Situation | Use || --- | --- || A topic passed `topic-precheck` (GO, GO (supporting content), or a NEEDS DECISION the reader settled) | this skill || Re-researching a topic before a rewrite | this skill, then `article-write` || The topic hasn't been prechecked | `topic-precheck` first || Auditing a published page against today's results | `refresh-audit` || Checking one draft's claims | `claim-check` | ## Inputs | Input | Required | Ways to provide, in order | If missing || --- | --- | --- | --- || The cleared topic | yes | a `topic-precheck` handoff; the reader's request | ask once; without a precheck, say it wasn't run and mark coverage `partial` || The top results for the primary keyword | yes | browse them; page text or headings the reader pastes; titles and URLs only | ask the reader to paste the top 5; titles only means matrix cells `not read` and coverage `partial` || The company's evidence | no | `## Claims and proof` in the profile; the precheck's evidence list; items the reader pastes | angle can't pass the non-commodity test; likely `COMMODITY` || Buyer questions | no | sales or support notes the reader shares; People Also Ask if browsing; community threads; `## Buying moments` | use the profile's moments and objections; coverage `partial` || Sources for figures | no | open each one in this run; source text the reader pastes | the figure goes to gaps | ## Profile read From `company-profile.md`: `## Company`, `## Product`, `## Buyers` (roles and objections), `## Buying moments`, `## Positioning` (unique attributes), `## Claims and proof` (with `### Never say`), `## Competitors` (mention policy), and every `## Learnings` entry for `article-research`. No profile: ask in one message for the buyer, what the product does, and any first-party proof on this topic. Say that `company-profile` saves this next time. ## Steps 1. **Apply learnings** for this skill (a source marked dead, an angle the reader rejected before). 2. **Check the precheck.** Read the handoff. `STOP` means `BLOCKED`: don't research a stopped topic unless the reader overrides it (record that as a learning). Carry forward the pillar, the page to avoid overlapping with, the evidence ids and the author. 3. **Read what ranks.** Follow `references/research-method.md` section 1 for 5 to 7 results. Record format, headings, core argument, data cited and weak spots per page. Say how many pages were read in full and how many were titles only. 4. **Build the coverage matrix.** Draft 6 to 12 subtopics (method section 2). Run:    ```   python3 scripts/coverage_matrix.py --input results.json --today <YYYY-MM-DD> [--fast] --ignore "<topic head term>"   ```    `--ignore` takes the topic's own head terms (they appear on every page, so matching on them marks everything covered). The script matches words, not meaning: read its matrix against the headings, correct any cell it got wrong, and say how many you changed. Without Python, mark each page per subtopic as `section` (a heading covers it), `mention` (the body touches it) or `none`, then class each subtopic: `baseline` if a section in 3 or more of 5 pages (60%: readers expect it), `thin` if mentioned but a section in at most 1, `gap` if no page covers it, `partial` otherwise. A page read as title only is `not read`, never `none`. Say which way you ran it. 5. **Find the freshest data.** For each claim the article will need, find a source that passes the credibility, recency and verification tests (method section 3). Open each one now. Run `claim-check` on the list, or apply its test inline. A figure that fails goes to gaps with what the source does say. Flag stats in the top pages that newer data replaces. 6. **Pull the company's evidence.** Map each item from `## Claims and proof`, the precheck list or the reader's paste to a subtopic. Strict relevance: the item must be about that subtopic, not share a word with it. Uncleared quotes and claims under "Claims to confirm" are listed as `needs clearance`, never used. 7. **Map buyer questions** (method section 4): the questions in the buyer's words, the vocabulary they use, the misconceptions, and which questions deserve their own section. 8. **Rank differentiation candidates** (method section 5): angle gaps, then depth gaps, then freshness gaps. Each needs evidence found in this run. 9. **Propose one angle.** One sentence on what this article says that the coverage matrix shows nothing ranking says, and one sentence naming the evidence that lets only this company say it. Apply the non-commodity test: would a careful summary of the top 5 contain it, or could a competitor publish it unchanged? If yes to either, it is not the angle. 10. **List E-E-A-T needs** (method section 6): experience, expertise, claims needing citation, and the YMYL flag. 11. **Suggest sections** in order: the active intent first, then the passive intents by what the reader needs next. These are suggestions for `article-write`'s outline, not the outline. ## Check and verdict | Verdict | Condition (first match wins) || --- | --- || `BLOCKED` | No topic, the precheck said STOP and the reader didn't override it, or no results could be read at all (not even titles) || `COMMODITY` | No candidate in step 8 is backed by evidence found in this run (the company's data, cleared customer words, a documented method or a first-hand finding), so no angle passes the non-commodity test: the article would be a rewrite of what ranks. Stop and route back to `topic-precheck` as a NEEDS DECISION || `READY WITH GAPS` | An angle passes the non-commodity test (provisionally, when pages were read as titles only), but some needs are open: pages not read in full, a subtopic with no evidence, a figure that failed verification, or a quote waiting for clearance || `READY` | An angle passes the non-commodity test, at least 5 pages were read in full, and every figure in the brief was verified in this run | Named stops: `BLOCKED` asks for the missing input. `COMMODITY` stops and tells the reader what evidence would change it (for example "a cleared customer example of <subtopic>" or "your own data on <figure>"). No other stops; `READY WITH GAPS` hands over with every gap labelled. ## Learn Append to `## Learnings` when: - the reader rejects the proposed angle (a `proposal` with their reason, such as "we don't publish anything on pricing");- a source turns out dead, wrong or paywalled (a `fact` naming the source, so later runs skip it);- the reader confirms a buyer question from sales or support (a `fact`, so `article-research` and `hook-writer` reuse it). Example: `- 2026-10-24 | article-research | fact | The 2019 attribution survey most top pages cite was replaced by a 2025 edition | Why: both sources opened 2026-10-24 | Status: noted` ## Output ```Verdict: <READY | READY WITH GAPS | COMMODITY | BLOCKED> · Coverage: <complete | partial (what's missing)>Topic: <topic> · Primary keyword: <keyword> · Researched: <date> · Pages read in full: <n> of <n> What ranks| # | Page | Date | Format | Core argument | Weak spot | Coverage matrix (by <script | reading>)| Subtopic | Page 1 | Page 2 | Page 3 | Page 4 | Page 5 | Class | Company evidence | Consensus: <what every top page agrees on>Where they differ: <the split, and which pages hold which side> Freshest data| Claim | Source | Date | Status (verified, failed, needs text) |Outdated figures in the top pages: <old figure, page> → <newer figure, source> Company evidence| Item | Type | Subtopic | Status (usable, needs clearance) | Buyer questions: <questions in their words, most important first>Vocabulary: they say <x>, not <y>Misconceptions: <buyers think x; the evidence says y> The angle: <one sentence> · Why only this company: <evidence ids>Other candidates: <type, one line each>E-E-A-T needs: experience <...> · expertise <...> · claims to cite <...> · YMYL <yes | no>Suggested sections: <ordered list>Gaps: <every open item, labelled>Learnings applied: <list, or none>``` Then the handoff block. ## Hand off ```handoffskill: article-researchversion: 1.0.0verdict: READY WITH GAPScoverage: partial (2 of 6 pages read as titles only)for: <topic>outputs:  primary_keyword: salesforce attribution reports  angle: <one sentence>  angle_evidence: [<claim or item ids>]  sections: [<suggested sections, in order>]  verified_sources: [<url>, <url>]  pillar: <URL or none>  author: <name, bio URL>  ymyl: falsenext: [article-write]open: [clearance for the customer quote on <subtopic>]``` For `COMMODITY`, `outputs.angle` is `not available`, `next` is `[topic-precheck]` and `open` names the evidence that would change it. ## Failure modes | Problem | Do this || --- | --- || The agent can't browse and the reader pasted titles only | Build the matrix from what was given, mark cells `not read`, and say the angle is provisional until the pages are read || A top result is a forum thread or a video | Record it as audience intelligence and its format; don't use it as a source || A competitor's page ranks in the top 5 | Read it like any other; follow `## Competitors` mention policy when the angle contrasts with it || The only evidence is a public study | It can support a section but not the angle; the angle needs something first-party || The reader's evidence contradicts a top source | Keep both, with sources, and say which the article should lead with and why || The topic turns out to be 2 intents | Say so, propose which one this article serves, and suggest the other for `topic-precheck` as its own topic | ## Influenced by - Google's guide to optimizing for generative AI features in Search, which separates non-commodity content (unique expert or experienced takes beyond common knowledge) from content that restates what is already available: https://developers.google.com/search/docs/fundamentals/ai-optimization-guide- SHAKEUP's content engine: the top-5 coverage matrix, evidence held to strict relevance, every source fetched in the same run, and "commodity" treated as a real finding that goes back to the precheck rather than a reason to write anyway.

How it thinks

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

What we built

In our own content engine, research ends with a coverage matrix of the top 5 and an angle backed by evidence found in the same run. Commodity is a real finding that goes back to the precheck, not a reason to write anyway.

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

    Check the precheck

    A stopped topic isn't researched unless you overrule it. The pillar, the page to avoid repeating and the evidence list carry forward.

  2. 02

    Read what ranks

    5 to 7 top results, read in full, with each page's headings, argument, data and weak spots. A page seen only as a title is marked as not read.

  3. 03

    Build the coverage matrix

    Its script marks each subtopic as a section, a mention or missing on every page. Baseline is what 3 of 5 cover. Thin and missing subtopics are where an angle can live.

  4. 04

    Check every figure

    Each statistic is checked against its source in this run. Old figures the top pages still use are flagged with the newer one beside them.

  5. 05

    Map your evidence and the buyer's words

    Your evidence is matched to subtopics by meaning, not by a shared word. Buyer questions are kept in their own phrasing, with the words they use.

  6. 06

    Propose one angle

    One sentence on what nothing ranking says, and the evidence that lets only you say it. If a summary of the top 5 would contain it, it isn't the angle.

What you need

  • Required

    The cleared topic

    A topic precheck handoff → Your request

  • Required

    The top results for the keyword

    It reads them itself if your agent can browse → Page text or headings you paste → Titles and URLs only

  • Optional

    Your own evidence

    The Claims and proof section of your company profile → The evidence list from the precheck → Cleared quotes, data or cases you paste

  • Optional

    The questions buyers ask

    Sales or support notes you share → People Also Ask, if your agent can browse → The buying moments in your profile

What it can return

  • READY

    An angle passes the test, 5 or more pages were read in full and every figure was checked.

  • READY WITH GAPS

    An angle passes, but something is open, such as pages read only as titles or a quote waiting for approval.

  • COMMODITY

    None of your evidence backs a different answer, so the article would rewrite what ranks. It goes back to the precheck.

  • BLOCKED

    There's no topic, the precheck said stop, or no result could be read at all.

Where it stops for you

  • When the topic turns out to be commodity, with the evidence that would change it

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 topic is attribution reports in Salesforce.

Verdict

READY WITH GAPS

Coverage: partial (pages known from headings and summaries, not read in full)

SubtopicPages that cover itClassOur evidence
Setting up Campaign Influence4 of 5, as a sectionBaselinenone needed
Attribution models4 of 5, as a sectionBaselinenone needed
Keeping opportunity stages intact1, in passingThinA product fact: stages are read, never rewritten
Setup time on SalesforcenoneGapOnboarding data: median 16 days on Salesforce, 12 on HubSpot
Freshest data

the top page cites a 2019 trust figure. A 2025 survey replaces it, checked against its text.

The angle

pipeline by channel in Salesforce without rewriting opportunity stages, and what that setup takes. Only this company can say it, because the setup figures come from its own onboarding.

The handoff block the next skill reads
skill: article-research
version: 1.0.0
verdict: READY WITH GAPS
coverage: partial (pages known from headings and summaries)
for: salesforce marketing attribution
outputs:
  angle: pipeline by channel in Salesforce without rewriting opportunity stages
  angle_evidence: [onboarding-salesforce-setup, doc-campaign-influence]
  pillar: /guides/b2b-attribution
next: [article-write]
open: [read the top 5 in full]

What good looks like

  • A matrix that separates what every page covers from what none of them do
  • Every figure checked against its source in this run, with old ones flagged
  • An angle that names the evidence only your company has
  • Buyer questions in the buyer's own words
  • Gaps left open and labelled, never filled

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.

Why a coverage matrix?

It shows what every top page already covers, which you must match, and what none of them cover, which is where an angle can live. Without it, an angle is a guess.

What makes an angle pass?

It rests on something only you have, such as your data, cleared customer words or your documented method. A competitor shouldn't be able to publish it unchanged.

Can it work without browsing?

Yes. Paste the top pages' text or headings. With titles only, it says the angle is provisional and asks for the pages.