tweet-insight

Tweet Insight

Starting from a tweet, deeply study all related content, then write an original sharing post in your own words.

Core Philosophy

This is not translation, not a summary, but "learning and sharing." It's like telling a friend what you learned after reading a paper - you've digested the content, formed your own understanding, and explained it in your own way.

Specific steps:

Read the original tweet → read the quoted tweet → read the related article/paper → organize the information → determine the topic and angle → write it up → review

Workflow

Step 1: Collect the tweet

The user provides one or more tweet URLs.

Preferred approach: Playwright MCP(Recommended, can retrieve complete structured content)

Use browser_navigate to open the tweet URL
Use browser_snapshot to get a page snapshot, and extract from it:
- Full tweet text
- Author, posting time, engagement metrics (replies, reposts, likes, bookmarks, views)
- Quoted tweet block - Note: the full URL of the quoted tweet does not appear directly in the snapshot
If there is a quoted tweet: click the link element of the quoted tweet block, the page will jump to the quoted tweet page; at this point get the full link from the Page URL, then use browser_snapshot to get the full content of the quoted tweet
Extract all external links from the tweet and the quoted tweet (t.co the actual URL that the short link points to)

Alternative approach: WebFetch + r.jina.ai(Use when Playwright is unavailable)

Use WebFetch(URL with the https://r.jina.ai/ prefix) to fetch the tweet content
Note: r.jina.ai can retrieve the tweet text and fragments of the quoted tweet content, but may not return the full URL of the quoted tweet
If the quoted tweet URL is missing, ask the user to provide it manually

Step 2: Deep-read the related content

Tweets often link to more important content. This step is key.

Extract all external links from the tweet (papers, blogs, announcement pages, system cards, etc.)
t.co Short links must be resolved before use: use Playwright browser_navigate to follow the redirect, and get the full real URL from the Page URL after the redirect. The link text displayed in the tweet snapshot (such as github.com/unslothai/unsl …) is truncated and must never be used directly in the output
Use WebFetch (optionally with the r.jina.ai prefix) to fetch the content of each link one by one
For PDFs: if WebFetch cannot retrieve them or the content is too large, prompt the user to download it and provide a local path, then use Read to read it
For long documents: focus on the abstract, conclusion, key data, and surprising details; no need to recount word for word
Valuable links should be collected: if the related content contains resource links useful to readers (guides, GitHub, notebooks, etc.), record their full URLs and include them in the final output

Step 3: Organize the information

After reading all the materials, sort out:

What is the essence of this matter? - Summarize the core event or finding in one sentence
Key data and facts - Extract the most impactful numbers and comparisons
Surprising details - The parts that make you go "wait, really?"
Practical impact - What this means for ordinary people/developers/the industry
Context and background - Why this is happening now, and how it differs from before

Step 4: Determine the angle and write the piece

Don't try to cover all the information. Pick one angle to切入 and unfold it like telling a story.

Writing principles

Hit the point right at the start - The first sentence should let people know what happened and make them want to keep reading
Speak with concrete data - Don't say "a significant improvement"; say "jumped from 42% to 97%"
Details are the soul - Thrilling, interesting, counterintuitive details give the content flesh and blood (for example, "the researcher was eating a sandwich in the park at the time")
Convey structure through content, not numbering - You can use paragraphs and blank lines to organize it, but avoid textbook-style "one, two, three"
Speak plainly - Write for a smart friend who may not necessarily know this field, not for an expert
Appropriately add your own judgment - Whether this thing is good, important, or credible, readers want to hear what you think

Format

Default output is a single tweet
Put the quoted link after the opening text - First use one or two sentences to introduce it, then place the link (this is both natural and generates a card); don't put it at the very beginning (too abrupt) or at the very end (it won't generate a card)
Structure depends on the content - Don't always use the fixed 1/ 2/ 3/ 4/ pattern. Some content suits bullet points, some suits continuous narrative, and some suits telling a story first and then giving data. Let the content itself determine the form
Primarily Chinese content (unless the original tweet is in English and the user has not specified a language)
Keep data and proper nouns in their original English
If the content is indeed very long, suggest that the user consider posting an X Article

Step 5: Review

After writing, self-check:

Factual accuracy - Can every data point be traced back to the original material? Are there no numbers made up from impression?
No overinterpretation - Where the original text says "may", you didn't write it as "definitely"?
No key information omitted - Are the most important 2-3 points covered?
Natural tone - Does it not read like machine translation or a press release?

If problems are found, correct them before presenting to the user.

Step 6: Apply personal writing style

Call the personal-chinese-writing-style skill to do a final round of polishing on the generated text:

Punctuation (use -for dashes, use ......for ellipses, use ""）
Article structure (structure hidden within the text, connected with prose rather than hard cuts)
Overall language feel

Step 7: Present the result

Output the final text, along with:

The original tweet link (as a reference source)
If the content is split into a tweet thread, note the word count of each tweet
Ask the user: use it directly, or adjust it?

Multiple tweets scenario

When the user provides multiple tweets:

First read and study them one by one
Determine the relationship between these tweets: are they different angles on the same topic? Or different topics?
Same topic → merge into one comprehensive sharing post
Different topics → generate them separately, or ask the user how they want to handle it

Notes

Never publish automatically, only output text
Do not simply translate the original tweets - if the original tweets are in English, your output should not read like a translation
Preserve the accuracy of the original data and terminology, but reorganize it in your own words
If there are claims in the original material that you cannot verify, flag them so the user knows