remorses/x-articles

<!-- Skill for editing x.com (Twitter) article drafts through Playwriter plus content-state mutations. -->

Use this skill when editing long-form article drafts on x.com/compose/articles (Twitter Articles).

Read Playwriter First

Before using this skill, read the playwriter skill and run:

playwriter skill

This skill assumes Playwriter is already set up and connected to the user's existing Chrome session.

Read the full output. Do not pipe it through head, tail, or other truncation commands.

Core idea

Use Playwriter for three things:

1. connect to the already-open x.com (Twitter) article draft
2. inspect the editor and capture one real network mutation
3. validate the final rendered result after updates

For anything bigger than a tiny tweak, do not rely on manual typing inside the editor. Generate the article content_state locally and send the same GraphQL mutation x.com (Twitter) already uses.

Editor model

The article body is represented as a content_state object with two main parts:

• blocks: ordered content blocks
• entity_map: supporting entities, especially code blocks

Important block types:

• unstyled — normal paragraph
• header-two — section subheading
• ordered-list-item — numbered list item
• atomic — embedded block like a markdown code block

Important entity type:

• MARKDOWN — used for code blocks, with the markdown fence stored in entity_map[*].value.data.markdown

Longer example content_state:

{
  "blocks": [
    {
      "key": "k0",
      "text": "event sourcing for application state",
      "type": "header-two",
      "data": {},
      "entity_ranges": [],
      "inline_style_ranges": []
    },
    {
      "key": "k1",
      "text": "your clanker loves state",
      "type": "unstyled",
      "data": {},
      "entity_ranges": [],
      "inline_style_ranges": [
        { "offset": 19, "length": 5, "style": "Bold" }
      ]
    },
    {
      "key": "k2",
      "text": "doubles your final app state",
      "type": "ordered-list-item",
      "data": {},
      "entity_ranges": [],
      "inline_style_ranges": []
    },
    {
      "key": "k3",
      "text": "doubles your bugs",
      "type": "ordered-list-item",
      "data": {},
      "entity_ranges": [],
      "inline_style_ranges": []
    },
    {
      "key": "k4",
      "text": " ",
      "type": "atomic",
      "data": {},
      "entity_ranges": [
        { "key": 0, "offset": 0, "length": 1 }
      ],
      "inline_style_ranges": []
    },
    {
      "key": "k5",
      "text": "if you can derive it, don't store it.",
      "type": "unstyled",
      "data": {},
      "entity_ranges": [],
      "inline_style_ranges": [
        { "offset": 7, "length": 6, "style": "Bold" }
      ]
    }
  ],
  "entity_map": [
    {
      "key": "0",
      "value": {
        "type": "MARKDOWN",
        "mutability": "Mutable",
        "data": {
          "markdown": "```typescript\nfunction shouldShowFooter() {\n  return true\n}\n```"
        }
      }
    }
  ]
}

This is the minimum mental model:

• blocks is the article in order
• each paragraph, heading, and list item is a separate block
• code blocks are atomic blocks that point into entity_map
• inline bold lives in inline_style_ranges

Recommended workflow

1. Open or locate the draft

Find the existing article editor page in the connected browser. The URL format is:

https://x.com/compose/articles/edit/<article_id>

Always parse and keep the numeric article_id. The content mutation needs it.

Example Playwriter check:

playwriter session new
playwriter -s 1 -e '
state.page = context.pages().find((p) => {
  return p.url().includes("/compose/articles/edit/")
})
if (!state.page) {
  throw new Error("No article editor page found")
}
console.log(state.page.url())
'

2. Explore with small manual edits first

Use the UI to learn how the editor reacts before doing bulk updates. Good exploration tasks:

• add one paragraph
• convert one block to Sottotitolo
• insert one code block
• bold one word in one paragraph

After each change, inspect the rendered HTML with getCleanHTML().

Example validation command:

playwriter -s 1 -e '
state.page = context.pages().find((p) => {
  return p.url().includes("/compose/articles/edit/")
})
console.log(
  await getCleanHTML({
    locator: state.page.locator("[data-testid=\"composer\"]"),
    showDiffSinceLastCall: false,
  }),
)
'

3. Capture real network traffic

Watch GraphQL requests while making one tiny manual change. This gives you the exact mutation names and payload shapes used by the current x.com (Twitter) editor.

The two important mutations found in this session were:

• ArticleEntityUpdateTitle
• ArticleEntityUpdateContent

The content mutation URL looked like:

https://x.com/i/api/graphql/<queryId>/ArticleEntityUpdateContent

The exact queryId can change over time. Do not hardcode it blindly without first confirming it from a real request in the current session.

Example request logger:

playwriter -s 1 -e '
state.page = context.pages().find((p) => {
  return p.url().includes("/compose/articles/edit/")
})
state.requests = []
state.page.removeAllListeners("request")
state.page.on("request", (req) => {
  if (req.url().includes("ArticleEntity") || req.url().includes("graphql")) {
    state.requests.push({
      url: req.url(),
      method: req.method(),
      postData: req.postData(),
    })
  }
})
console.log(
  "Ready: now make one tiny manual edit in the page, then rerun this command to inspect state.requests",
)
'

4. Use direct content updates for bulk work

Once you know the current mutation shape, generate the full content_state locally and send the content update directly.

This is the reliable path for:

• full markdown import
• replacing large sections
• converting paragraphs to ordered lists
• adding one bold keyword per paragraph
• fixing code block languages

Concrete pattern:

1. build content_state in a local JSON file
2. read ct0 from document.cookie
3. send ArticleEntityUpdateContent with the same queryId and feature flags
4. reload the page

5. Reload and validate

After every direct mutation:

1. reload the article editor page
2. inspect getCleanHTML()
3. search for expected headings, list items, bold splits, and code labels

Do not trust the visual editor alone.

Example reload + search:

playwriter -s 1 -e '
state.page = context.pages().find((p) => {
  return p.url().includes("/compose/articles/edit/")
})
await state.page.reload({ waitUntil: "domcontentloaded" })
await waitForPageLoad({ page: state.page, timeout: 8000 })
console.log(
  await getCleanHTML({
    locator: state.page.locator("[data-testid=\"composer\"]"),
    search: /debugging with event streams|typescript|ordered-list-item/i,
    showDiffSinceLastCall: false,
  }),
)
'

Block type cheatsheet

Paragraphs

Use:

{
  "type": "unstyled",
  "text": "your paragraph text"
}

Subheadings

Use:

{
  "type": "header-two",
  "text": "debugging with event streams"
}

Numbered lists

Each item is its own block:

{
  "type": "ordered-list-item",
  "text": "doubles your bug surface"
}

Code blocks

Code blocks are not plain text blocks. They are:

1. one atomic block in blocks
2. one MARKDOWN entity in entity_map

The atomic block points to the entity with entity_ranges.

The entity markdown should include the full fence, for example:

```typescript
const x = 1
```

If you want the visible language label to say typescript, the stored fence must be ```typescript, not ```ts.

Inline styles

Bold text is represented with inlineStyleRanges inside a block.

Important session learning:

• the style name is Bold
• not BOLD

Example:

{
  "text": "your clanker loves state",
  "inlineStyleRanges": [
    { "offset": 19, "length": 5, "style": "Bold" }
  ]
}

Always calculate offsets against the raw block text exactly as stored.

Known UI pitfalls

The manual editor flow has several traps:

Heading inheritance

After creating a heading, pressing Enter once can keep the next block in the same heading style. To reset to a paragraph, press Enter again.

Post-code-block cursor placement

Typing after a code block is unreliable. The editor can:

• append text to the wrong block
• split text unexpectedly
• create stray headings
• leave part of a sentence in one block and the rest in another

For anything more than a tiny manual tweak, use direct content updates instead.

Visual feedback is incomplete

The editor can look correct while the underlying block structure is wrong. Always inspect the HTML or mutation payload.

Playwriter sessions can reset

If the relay server restarts or the extension reconnects, Playwriter sessions can disappear. If that happens, create a new Playwriter session and reattach to the already-open article page.

Recovery command:

playwriter session new
playwriter -s 1 -e '
state.page = context.pages().find((p) => {
  return p.url().includes("/compose/articles/edit/")
})
if (!state.page) {
  throw new Error("No article editor page found")
}
console.log(state.page.url())
'

Auth and request details

Direct content updates need proper auth headers. In this session, the direct fetch() worked only after including:

• the X bearer token
• x-csrf-token from the ct0 cookie
• the standard X active-user/auth/client-language headers

If you get 403, inspect the successful browser request and match its headers.

In this session, the direct fetch succeeded only after matching:

• bearer token
• x-csrf-token
• x-twitter-active-user
• x-twitter-auth-type
• x-twitter-client-language

Validation checklist

After updating an article, verify all of these:

1. correct title in the title field
2. headings appear as header-two
3. ordered lists appear as ordered-list-item
4. code blocks render as markdown-code-block
5. code block language labels say what you expect, for example typescript
6. bold keywords are split into separate styled spans in the HTML
7. no stray empty headings or broken split paragraphs remain

Useful recipes

Import a markdown article

1. parse the markdown locally
2. map paragraphs to unstyled
3. map ## headings to header-two
4. map numbered list items to ordered-list-item
5. map fenced code blocks to atomic + MARKDOWN entities
6. send ArticleEntityUpdateContent
7. reload and validate

The fastest implementation is usually:

1. generate ./tmp/x-article-content-state.json
2. read it from a Playwriter command with fs.readFileSync
3. push it with the direct content mutation

Bold one keyword per paragraph

1. choose one keyword per paragraph
2. compute exact offset and length
3. add inlineStyleRanges with style Bold
4. push the updated content_state
5. reload and verify the HTML splits around the bold span

Fix code language labels

Update the markdown entity fences. Example:

• bad: ```ts
• good: ```typescript

Then resend the full content_state and reload the editor.

Minimal bulk update example

Use this pattern when you already have the right queryId and payload shape:

playwriter -s 1 -e '
const fs = require("node:fs")
state.page = context.pages().find((p) => {
  return p.url().includes("/compose/articles/edit/")
})
const articleId = state.page.url().match(/edit\/(\d+)/)?.[1]
const contentState = JSON.parse(
  fs.readFileSync("./tmp/x-article-content-state.json", "utf8"),
)
const csrfToken = await state.page.evaluate(() => {
  return document.cookie
    .split("; ")
    .find((x) => x.startsWith("ct0="))
    ?.slice(4) || ""
})
const payload = {
  variables: {
    content_state: contentState,
    article_entity: articleId,
  },
  features: {
    profile_label_improvements_pcf_label_in_post_enabled: true,
    responsive_web_profile_redirect_enabled: false,
    rweb_tipjar_consumption_enabled: false,
    verified_phone_label_enabled: false,
    responsive_web_graphql_skip_user_profile_image_extensions_enabled: false,
    responsive_web_graphql_timeline_navigation_enabled: true,
  },
  queryId: "<capture-from-real-request>",
}
const response = await state.page.evaluate(async ({ payload, csrfToken }) => {
  const res = await fetch(
    `https://x.com/i/api/graphql/${payload.queryId}/ArticleEntityUpdateContent`,
    {
      method: "POST",
      credentials: "include",
      headers: {
        authorization: "<capture-from-real-request>",
        "content-type": "application/json",
        "x-csrf-token": csrfToken,
        "x-twitter-active-user": "yes",
        "x-twitter-auth-type": "OAuth2Session",
        "x-twitter-client-language": "it",
      },
      body: JSON.stringify(payload),
    },
  )
  return { status: res.status, text: await res.text() }
}, { payload, csrfToken })
console.log(response.status)
console.log(response.text.slice(0, 1000))
'

Replace the bearer token and queryId with values captured from a successful browser request in the current session.

Default strategy

Use this default unless the task is tiny:

1. inspect the current draft in the browser
2. capture one real content mutation from X
3. generate the final content_state locally
4. update the draft with the same mutation shape
5. validate the result in the live editor HTML

That is the fastest path and the most likely to work in one shot.