long-form-article
Plans, researches and drafts one long guide (3,000 to 5,000 words) for an X Article, a LinkedIn article, a blog or a newsletter, with a point a reader could repeat, named and dated sources, a proven structure from the size of the gap to "start this week", copyable prompts or steps, an honest "what breaks", and optional images. Use when you want a long-form piece, a playbook, a masterplan or a definitive guide that becomes the go-to resource on a topic.
How do I install this agent skill?
npx skills add https://github.com/boringmarketer/boring-marketing-skills --skill long-form-articleIs this agent skill safe to install?
- Gen Agent Trust Hubpass
This skill is a robust tool designed for researching and writing high-quality, long-form articles. It utilizes local Python scripts for draft validation and integrates with external research services. The analysis found no evidence of malicious behavior, data exfiltration, or harmful intent.
- Socketpass
No alerts
- Snykwarn
Risk: MEDIUM · 1 issue
What does this agent skill do?
Long-Form Article
This skill produces one long guide that a reader finishes, saves and acts on. Done means a draft where the point fits in one sentence, every number names who measured it and when, the reader can do something this week, the honest sections are there ("what breaks", the vendor-data caveat), and every line that needs the author's OK is flagged.
Length earns its place only when a reader can repeat the point, check every number and act on it. Without those three, 4,000 words is a blog post with padding. The structure below is the one that held up across real articles; it works for an opportunity playbook, a how-to for buyers, and a guide a local business writes for its customers. This is a floor, not a ceiling. When the piece in front of you argues for a better shape, take it and say why in one line.
Context
Only if ./brand/, ./brands/ or MARKETING.md exists (nothing to load otherwise), run:
python3 ${CLAUDE_SKILL_DIR}/../start-here/scripts/brand_context.py --for long-form-article
(${CLAUDE_SKILL_DIR} is this skill's folder; outside Claude Code use the folder this file was read from, e.g. .agents/skills/long-form-article. If the script or python3 is unavailable, read the brand folder directly: ./brands/<slug>/ when ./brands/active names one, else ./brand/; then MARKETING.md, preferences.md and the files in reads:. If none exist, proceed standalone and ask only for the point, the reader and the platform.)
voice-profile governs how it sounds; positioning#chosen-angle keeps the piece on the author's lane; audience#pain-language supplies the reader's own words for the title and the opening. Apply any [long-form-article] learnings (which titles, sections or links performed). One that changed the piece gets one line in the Sources and assumptions note, only if it did.
Modes
| Situation | Path |
|---|---|
| A topic or a point, nothing written (default) | Brief, research, outline, draft, gate, images, platform prep |
| A draft exists | Run the gate (Phase 5) on it, fix the weakest dimension first, then platform prep |
| "Just outline it" | Brief, research, outline. Stop. |
Phase 1: The brief
Pull answers from the conversation first. Ask only for gaps, all in one message, four questions at most.
- The point, in one plain sentence a reader could repeat. Test: say it to someone in the field. If they ask "so what do I do?", you have a topic, not a point. Three shapes that pass (invented examples):
- A SaaS founder: "Most clinic no-shows are a reminder problem, and three well-timed texts fix most of it."
- An agency owner or creator: "The easiest agency to start this year turns long podcasts into short clips, because every podcaster needs it and almost none have the time."
- A local business, Hartline Plumbing in Dallas (a sample business): "Replace a water heater before it fails, because a planned swap costs less than an emergency one and you get to pick the model."
- The reader, and what they can do after reading. One concrete action, this week.
- The platform. X Article, LinkedIn article, blog post or newsletter. It changes the build, not the writing.
- What the author allows about themselves. Which credentials, numbers and stories can be printed. By default, name no client.
- Anything to link. The author's product or offer, with what it really does, checked in its docs or code. Never describe a feature that doesn't exist.
- The author's past long pieces. Read them. The new one gets its own angle, its own title and its own structure. Reuse a device only when the new piece needs it, and never reuse a line.
The title names the reader's outcome in plain words. "How to..." works. A title that names the author's method instead of the reader's result reads as inside baseball. Offer three, recommend one.
Phase 2: Research
Fan out in parallel (one research subagent per source type, each with a small fetch and time budget; one WebSearch in the main thread first to confirm the web works):
- The numbers. Load
references/research-sources.md. Government data and large studies first. Vendor numbers carry "(vendor data)". Every number goes intosources.mdwith the source, its date, the sample, the URL and the exact figure. - The author's own material. Only files in this workspace or what the user hands over. This is where the voice and the first-hand stories come from. Never the assistant's memory or other folders.
- The pattern read. If the user names two or three pieces to model, read them for structure (what each part does, where the proof sits), never for sentences.
- Product facts, if the piece links to one.
A number without a source gets cut or marked [confirm]. Specific beats round: "2,412 businesses" beats "thousands", and always says who counted.
We use Exa to find the studies and primary sources a keyword search misses (free tier, then pay as you go), and Firecrawl to read each one in full so every figure is checked at the source (free tier of 1,000 credits a month, no card). Without them, WebSearch and WebFetch do the job. Open the source page for every figure rather than trusting a snippet.
Phase 3: The outline (settle the shape before drafting)
Load references/structure.md. It holds the twelve parts, why each one works, how each part changes for a playbook, a how-to or a customer guide, and short invented examples across all three.
The parts, in order: the point and the read time · the size of the gap · the whole idea in four steps · how we got here (eras) · why this audience or market · what to offer · the money math, labeled "arithmetic, not a promise" · how to do it, with copyable prompts or steps · keeping it working · what breaks · start this week · the close.
Write outline.md: the title options, every H2, one line per section on its job, where each sourced number goes, where the author's credential sits (mid-article, inside a point it supports), and the image slots. Show the outline and the recommended title in one message and get one OK before drafting. One settled outline saves more rounds than any edit pass.
Phase 4: Draft
Write to article.md, section by section, from the outline. The rules (load references/writing-rules.md for the full list with examples):
- No em dashes or en dashes. Commas and periods.
- Sources as plain names in parentheses, never links: "(Census, 2022 Economic Census)", "(Pew Research, 2025)", "(Acme Survey, 2026, vendor data)".
- Credentials mid-article, one line, inside a point they prove. Never in the opening.
- No income promises. Money math is labeled "arithmetic, not a promise" and sends readers to their own numbers. A clearly labeled hypothetical worked example is fine.
- Real lists. Numbered steps and bold-led items, rendered by the editor, not "1)" typed into a paragraph.
- Copyable prompts or steps in code blocks, with
[placeholders]in square brackets. Every prompt follows the article's own advice and gets run once before the draft is called final. - Links where the reader would act, one per paragraph at most, and a call to action on its own short line. A link at the end of a long paragraph doesn't get clicked. Give every link its own
utm_contentso the next-day read shows which one worked. - If the author sells a tool, be honest about which steps it does. Give away prompts for the work it doesn't do; for the work it does, say so plainly.
- Leave out "hot take" labels, emojis, follow-me plugs unless asked, a moral as the last line, and legal acronym dumps (keep the plain rule: "you can't robocall a cell phone without consent", not the statute).
Read time: words divided by about 240 (Brysbaert, 2019, a meta-analysis of reading-rate studies), rounded, stated in the opening.
First drafts run long. One article was cut by almost half and got better. Cut sections that fail their deletion test before cutting sentences.
Phase 5: The gate
Run the mechanical scan first: python3 ${CLAUDE_SKILL_DIR}/scripts/scan.py article.md. It reports word count and read time, every dash, every leftover placeholder ([confirm], TODO, TK, $X), paragraphs with a number and no source in parentheses, paragraphs with more than one link, and the H2 list for the skim test. Fix what it finds.
Then score six dimensions, 1 to 5, as a silent internal check. The author sees only the passing piece:
| Dimension | 5 looks like |
|---|---|
| Point | A reader could repeat it after the opening, and every section serves it |
| Proof | Every number has a named, dated source; vendor data is labeled; first-hand counts are real |
| Use | The reader can act this week; the prompts and steps were run once and work |
| Honesty | "What breaks" is real, nothing is promised, the vendor-data caveat is there |
| Plain | Reads like one person explaining it; none of the tells in references/writing-rules.md |
| Shape | A skimmer gets the whole argument from the H2s and the four-step list |
Ship at 25 of 30 with nothing under 4. Below that, rewrite the weakest dimension and rescore, up to three passes. If a weakness survives the third pass, say what it is in one line in the Sources and assumptions note. No scores or score history are shown.
The fresh-eyes pass. Give a second agent only the brief and the draft, and ask what a stranger would trip on, where a claim outruns its source, and which link or prompt overpromises. On one long article it returned more than a dozen fixes.
Every [confirm] line goes to the author in one list. None stay in the file the author publishes.
Phase 6: Images (optional)
Skip for a blog or newsletter that doesn't need them. For X and LinkedIn, the cover decides the click. Load references/images.md: cover rules (one subject, five words or fewer, readable at feed size, adds a hook rather than repeating the title), one body image per major idea every 400 to 600 words, exact text as code blocks rather than generated images, and a check of every word and number in every image. Generate through /creative.
Phase 7: Platform prep
Load references/platform-notes.md for the platform's dated facts (who can publish, cover sizes, what the editor drops, email clipping). For an X Article built in the browser, references/x-articles-build.md is optional reading.
Hard rule: never publish. This skill ends at a finished draft, or a saved, unpublished draft in the platform's editor. The author publishes.
Output
./campaigns/{slug}/article.md is the deliverable, ready to paste: title, read time, body, image slots with captions, and a sources list at the end for the author (not for publishing). Beside it: outline.md, sources.md, and images/ if made.
## LONG-FORM ARTICLE: {Title}
{date} · a {platform} article for {the reader}
**The point:** {one sentence}
**The piece:** {n} words, about {n} minutes to read, ready to paste.
**For the author to confirm:** {every [confirm] line}
**Sources and assumptions:** {10 lines at most}
**Files saved:** ...
**What's next:** ...
Files first, then the chat summary, which never pastes the article. The Sources and assumptions note says in plain words what the numbers rest on (government and study data, vendor data, the author's own counts, each with how many), what the author must confirm, and what is missing. The gap-aware receipt is at most 3 of those lines: what was checked at the source, what could not be, and the one input that would unlock the next step (for example "the three first-hand counts need your real numbers"). No score, no data tag, no tool names. Append one line to ./brand/assets.md if a brand folder exists.
When the user edits your work
Apply the edit, then make it stick: save a gold case to ./brand/gold/<this skill>/ and
one to three yes/no checks to ./brand/checks.md, and confirm in one line. Before handing
over any work, answer the checks the context block loaded, silently, and fix every no.
How: ${CLAUDE_SKILL_DIR}/../start-here/references/taste-loop.md.
What's next
/content-atomizer: turn the article into posts for every platform the same day (~15 min)/social-content: the launch post and a quote-post image of the core framework (~10 min)/newsletter: send it, or a short version with the link, to the list (~15 min)/seo-content: when the blog version needs to rank, the search pass (~20 min)
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/boringmarketer/boring-marketing-skills/long-form-article">View long-form-article on skillZs</a>