This is a guide to writing a long-form article on X, built on 296 articles from 12 accounts.
An X article is X’s long-form format: a post with a title, a cover image, headings and images inside, read on its own page. The accounts in the data write about go-to-market (GTM), outbound and growth. GTM means everything a company does to find and win customers.
The data is my own. I collected the articles, described each one on 25 fields with Claude Code and compared every article only with other articles by the same author.
The clearest result: articles that open with a result beat their author’s typical article in 6 of the 7 accounts that wrote enough of them to test. Articles that open with a hot take, a provocative opinion, did it in 1 of 6.
This guide is for founders who want to grow on X with long-form articles, and for operators who already write them. You get ten rules with their numbers and a checklist.
If you only want the rules, read the next section and the checklist at the end.
The running example: say you run sales at a US software startup and you’ve drafted “9 Cold Calling Tips That Work”. It’s 700 words, has no images and opens with “Everyone says cold calling is dead.” Each rule below changes one thing about it.
The ten rules
- Pick a format that covers a whole thing. Full courses and company breakdowns rank highest. Tip lists and opinion essays rank near the bottom.
- Write for one reader. Articles written for everyone ranked lowest of all audiences.
- Title it as a breakdown, a complete guide or a curiosity. Numbered titles like “9 ways to …” come last.
- Open with a result or a promise. Result openers beat the author’s typical article in 6 of 7 accounts.
- Write 1,000 to 3,000 words. Under 1,000 is the one length that clearly loses.
- Write short paragraphs in plain English. Articles with paragraphs under 20 words beat the typical article in 6 of 7 accounts.
- Be specific and name real companies. Reach rises with every step of specificity.
- Use four or more designed images. Articles with 4 or more images beat the typical article in 4 of 4 accounts. No image at all: 0 of 8.
- Put tool logos and a short title on the cover. Logo covers beat the typical article in 4 of 4 accounts.
- End with one keyword ask for a free resource. This is the weakest rule, but the direction holds.
How the analysis works
Collecting. I collected 296 long-form articles from 12 accounts that write about GTM, outbound and growth. For each one I kept the full text, every image and the public counts of views and bookmarks. A bookmark is a reader saving the article to read later. The articles hold 884 images in total.
Describing. Every article got the same 25 fields: topic, format, title type, opener type, kinds of proof, the part a reader can copy, the ask at the end, and scores from 1 to 5 for things like specificity and readability.
Most fields have a fixed list of allowed answers, so 296 answers can be counted.
Claude Code did the reading. Claude Code is Anthropic’s AI assistant that works with the files on your computer. Twelve Claude Code agents, separate sessions that each work through their own batch, wrote one file per article. A pilot on 6 articles ran first and showed four fields that were missing.
The agents never saw the view counts, on purpose. I wanted each article described without a guess about why it did well.
Counting. A Python script joined the fields with the numbers and did all the math. No AI model calculated a number.
Accounts with fewer than 5 articles and articles younger than 7 days stayed out of the performance numbers. A rank among three articles means little, and a four-day-old article keeps collecting views for a while. That leaves 275 articles from 8 accounts.
Images. Later every cover (294, two articles had none) and every one of the 884 images inside the articles got its own fields: the type of image, the number of logos, the words on the cover, designed graphic or screenshot.
Compare each article with its own author
Views mostly measure account size. In this data, the weakest article of the biggest account got 16,241 views. The best article of a smaller account got 815. If you sort 296 articles by views, you mostly learn who has the most followers.
So every article is ranked against its author’s other articles. Its rank is where its views fall among that author’s articles, from 0 to 1. A rank of 0.50 is the author’s typical article, and from 0.75 up, an article is among the author’s best 25%.
When a group of articles averages 0.59, those articles did better than their authors usually do. Every number below is such an average rank.
Then count accounts. One account wrote 125 of the 296 articles, and a rule that only holds there is weak. So for every rule I count in how many accounts the group beats the author’s typical article.
An account only counts if it wrote at least two such articles. “6 of 7” means the rule held in six of the seven accounts where it could be tested.
Don’t judge by bookmark rate. Bookmark rate is bookmarks divided by views, and it drops when an article reaches far beyond the author’s followers. My explanation: people who don’t follow you bookmark less often than people who do.
Full courses show how far this misleads. They rank 0.17 on bookmark rate and 0.81 on the number of bookmarks. Template collections go the other way: 0.72 on rate and 0.35 on bookmarks.
So for bookmarks I use the number of bookmarks, ranked within the author.
Two limits. These are correlations, so a rule shows what goes together with reach, and the cause can be something else. Some groups are small, and I’ll say so where they are.
Run the same comparison on your own articles. Once you have ten or more, put the title, date, views and bookmarks of each one into a CSV file, a plain spreadsheet file. Add two columns by hand: the format and the type of opener. Then paste this into Claude Code:
Read my_articles.csv (title, date, views, bookmarks, format, opener).
Leave out articles younger than 7 days.
For every article, compute its rank among my own articles,
from 0 to 1, once for views and once for bookmarks.
Then print the average rank per format and per opener,
with the number of articles in each group.
Do all math in a Python script. Never estimate a number.
1. Pick a format that covers a whole thing
Format separates the articles more than almost any other field. From the top:
- Full course: 0.88, and every one with performance data was among its author’s best 25%. Only 4 full courses had performance data, and only one account wrote two or more, so treat it as a strong hint.
- Company breakdown: 0.86. 5 of 7 were among the author’s best 25%, and it beat the typical article in 2 of 2 accounts.
- Tool stack: 0.58.
- Tutorial or build-along: 0.55, in 3 of 4 accounts.
- Step-by-step playbook: 0.53, in 4 of 6 accounts.
At the bottom sit case studies at 0.47 (0 of 4), frameworks at 0.46 (0 of 3) and tip lists at 0.38 (0 of 3, none among an author’s best 25%). Opinion essays reach 0.34 (1 of 5) and template collections 0.30 (0 of 2).
My read: the formats at the top hand the reader a complete thing. The formats at the bottom hand over pieces the reader has to put together.
In the example, the nine loose tips become one playbook for a calling block, a fixed hour in which the reps only make calls. It covers every step in order, up to the booked meeting.
2. Write for one reader
Articles written for a general audience averaged 0.21, the lowest of any audience. The group is small (13 articles, 7 with performance data, testable in 2 accounts), but the gap is large. Articles for founders, agency owners, GTM engineers and marketers all landed between 0.50 and 0.54.
In the example: the reader is a manager with three sales reps who book first meetings by phone. That choice decides every number in the article.
3. Title it as a breakdown, a guide or a curiosity
Titles follow the formats. Breakdown titles average 0.71. Complete guide and full course titles average 0.69 and beat the typical article in 3 of 3 accounts. Curiosity titles, which hint at something without saying what it is, average 0.62.
A result in the title (“How I got X”) averages 0.52 and holds in 5 of 7 accounts. A plain “How to” without a result sits at 0.48, contrarian titles at 0.41, and numbered titles come last at 0.30, in 0 of 3 accounts.
Two smaller signals: a dollar amount in the title beat the typical article in 5 of 7 accounts, and so did a title in the first person.
In the example: “How to Run a Cold Calling Block: The Complete Playbook”.
4. Open with a result or a promise
Result openers and promise openers both average 0.59. Result openers beat the typical article in 6 of 7 accounts, promise openers in 4 of 5.
Hot takes average 0.43 and beat the typical article in 1 of 6. Openers that start with the reader’s pain rank lowest at 0.41, in 0 of 4.
My read, and the data can’t prove it: this is the opposite of short posts, where a hot take makes people stop scrolling.
Someone who opens an article has already clicked. They want to know what they get for the next ten minutes, and a hot take delays that answer.
The shape I use for the first lines: say what the article is (“This is a full course on …”). Then give the running example in one sentence. Then give the main result in exact numbers and say who the article is for.
In the example, “Everyone says cold calling is dead” becomes “This is the full cold calling block for a team of three reps, with the opener script and the order to call in.”
5. Write 1,000 to 3,000 words
Under 1,000 words, articles average 0.38 and beat the typical article in 2 of 7 accounts. From 1,000 to 1,999 words they average 0.55 (5 of 6), from 2,000 to 2,999 they average 0.52 (3 of 5), and at 3,000 words or more they average 0.53.
Past 1,000 words, length barely moves anything. Short is what loses.
In the example, the 700-word draft grows to 1,800 words, because each step now carries its script and what to do when it fails.
6. Write short paragraphs in plain English
Two fields point the same way. Articles whose paragraphs average under 20 words rank 0.57 and beat the typical article in 6 of 7 accounts. Articles with paragraphs of 30 words or more rank 0.48, in 0 of 5.
The agents also rated how easy each article is to read, from 1 to 5. The 20 articles rated 3 averaged 0.33 and beat the typical article in 0 of 4 accounts. Articles rated 4 or 5 averaged 0.51.
I write for a smart reader whose English is good but not native, and who has never done the task. Five habits help that reader most:
- No idioms. “Boil the ocean” becomes “call every company in the state”.
- Explain every term the first time, in one short clause. Write “TAM (total addressable market)” the first time, and “TAM” after that.
- One word, one meaning. If “lead” means a person in step 2, it can’t mean a company in step 5.
- Headings say what to do. “Write the opener before you build the call list” works. “Openers” doesn’t.
- Give every claim its context. Each number or tip needs one sentence before it that says why it’s there.
Links make no difference either way: 152 of the 296 articles have none, and they do as well as the rest. So I leave them out of the body.
In the example, the tip “Dial smart” becomes the heading “Call the accounts that answered last week first”.
7. Be specific and name real companies
Every article got a specificity score from 1 (vague advice) to 5 (numbers, tools, settings, real steps). Reach climbs with it: 0.32 at a score of 2, 0.45 at 3, 0.53 at 4 and 0.55 at 5. Articles with no proof of any kind averaged 0.32.
Naming real companies was the strongest single kind of proof, at 0.65 and above the typical article in 5 of 5 accounts. It fits the breakdown result: readers want to see the thing working at a company they know.
Four habits make an article specific:
- Write measured numbers exactly. If you counted 2,317, write 2,317, never a rounded 2,300. Numbers in an example may stay round, as long as the text says it’s an example (“say your list has 40,000 companies”).
- Name two to four tools per job, the way the market names them. If one charges a flat monthly fee and another charges per result, say so, because that decides which one to use first. One tool from your own setup reads like an ad.
- Explain every decision a reader might question, in one sentence. If you kept companies without a website, say why.
- Keep only numbers that teach the reader something. How many days your project took rarely helps them. What to expect from a step does.
In the example, “keep your opener short” turns into the three-line opener, word for word, and the tips name the dialer and the CRM fields they use.
8. Use four or more designed images
Articles with 4 or more images average 0.59 and beat the typical article in 4 of 4 accounts. One to three images: 0.47, in 3 of 5. No image at all: 0.44, in 0 of 8 accounts. That last group is big: 106 of the 296 articles had no image.
The kind of image matters too. For images and covers I used a combined rank, the average of the views rank and the bookmark rank, both within the author. 0.50 is still the typical article. The groups are small, so read this as a direction:
- Text cards go with weaker articles. A text card is an image that is mostly text on a background. The 35 articles with one averaged 0.41, against 0.56 for articles with images and no text card. Give a checklist icons and short lines instead.
- Real tool logos help. The 30 articles with an image that shows two or more logos averaged 0.60, against 0.51 for the other articles with images.
- Mix in one real screenshot. The 15 articles that mixed designed graphics with at least one real screenshot averaged 0.64.
In the example: a flow of the calling block, a table of call outcomes, a chart of calls per booked meeting and a real screenshot of the dialer view.
9. Put tool logos and a short title on the cover
The cover is the image people see in the feed before they click. 294 of the 296 articles had one, and these numbers use the same combined rank as rule 8.
- Logo covers rank highest. Covers where tool logos are the main picture average 0.63 and beat the typical article in 4 of 4 accounts (24 articles).
- Use 8 words or fewer. Covers with 1 to 8 words average 0.53, in 5 of 6 accounts. With 9 to 20 words they average 0.42. With 21 or more words, 0.46, in 0 of 5.
- Leave out the big number, the screenshot and the AI picture. Covers built around one big number average 0.46 (1 of 6 accounts), app screenshots 0.43 and AI-generated art 0.42.
- Color makes no difference. Dark and light covers both average 0.50, and so do purple ones. Pick one look and keep it.
In the example: two logo tiles, the dialer and the CRM, with the title “How to Run a Cold Calling Block” and “Complete Playbook” in handwriting above it.
10. End with one keyword ask
Articles that asked readers to comment or send a direct message with a keyword, usually to start a sales conversation, averaged 0.60. A free resource behind a link, comment or message averaged 0.54. A product plug averaged 0.45, and no ask at all 0.48.
This is the weakest rule here. Per account the picture is mixed: the comment keyword held in 2 of 4 accounts, and product plugs beat the typical article in 3 of 6.
How much an article promotes a product barely mattered either. Articles scored from 0 to 3 for promotion all landed between 0.49 and 0.51.
My default is one keyword for a free resource that matches the article, at the very end, as one line: “Comment CALLS and I’ll send you the script.” It’s easy for the reader, and the direction of the data agrees with it.
In the example: the keyword CALLS, and the full script as the resource.
The checklist
Run your next draft through this before you publish:
FORMAT, READER, TITLE
[ ] The article covers a whole thing: a course, a company, a build or a playbook
[ ] Written for one reader you could describe in one sentence
[ ] The title is a breakdown, a complete guide, a result or a curiosity
[ ] No numbered title ("9 ways to ...")
OPENER
[ ] Line 1 says what the article is or states a result
[ ] No hot take, no pain point, no question in the first three lines
[ ] The main result in exact numbers, then who it's for
BODY
[ ] 1,000 to 3,000 words
[ ] Short paragraphs of one to three sentences
[ ] Every term explained the first time, no idioms
[ ] Every heading says what to do
[ ] Measured numbers exact, example numbers framed as examples
[ ] At least one real, named company
[ ] Two to four tools named per job
IMAGES
[ ] 4 or more designed images, no text cards
[ ] Real tool logos where a tool is named
[ ] One real screenshot
[ ] Cover: tool logos, 8 words or fewer, no big number
ENDING
[ ] One ask, at the very end
[ ] A keyword for a free resource that matches the article
To check a draft against the list, paste the checklist and your draft into Claude and add: “Go through every line of the checklist. For each one, answer pass or fail and quote the sentence from my draft that decides it. Don’t rewrite anything.”
Where to start today
You can apply the ten rules to your next article in three steps:
- Run your last article through the checklist and mark every rule it misses.
- Fix the opener, the cover and the image count first. They are the fastest to change and among the most consistent rules in the data.
- Write your next article to the checklist and compare its views and bookmarks with your own earlier articles, with the prompt from the section on comparing, not with other accounts.


