SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsx-post-research

x-post-research

Writing & research

View repository

Updated 10-15 16:46

About this skill

X Post Research

A workspace skill that collects public posts starting from the FxTwitter API and organizes primary-source URLs, related GitHub repos, images, and discussion points under research. It can start not only from hashtags but also from ordinary keywords or known post URLs.

When to Use

  • When you want to pick up reactions, third-party evaluations, and real-world usage reports after the announcement of a new model or new feature name (e.g.: "GPT-6 Luna" OR "GPT6 Luna")
  • When you want to look for new technologies or use cases from X ordinary keywords,OR, domains, or known post URLs
  • When you want to follow announcements on the day of an event from X hashtags
  • When you want to extract posts with many reactions as candidates and check them through to the primary deliverables
  • #MSBuild や #MicrosoftBuild When you want to sweep through event hashtags such as in the last few hours
  • When you want to gather primary-source URLs and repo leads first, right after a keynote
  • When you want to converge from noisy live commentary posts to Microsoft Learn / blog / GitHub / session repo
  • When you also want to keep posts with images as research assets

Scope

  • This skill is dedicated to X
  • Bluesky, LinkedIn, YouTube comments, etc. are out of scope
  • Posting to X and account operations are out of scope
  • If you want to handle those, split them out into a separate skill

Default Profile

This is the default candidate when no explicit specification is given.

  • time window: 直近 4 時間
  • target posts: 500
  • raw artifact: tmp/
  • research note: research/YYYYMMDD-<topic>-from-x.md
  • curated images: research/assets/YYYYMMDD-<topic>-from-x/
  • bulk images: research/images/YYYYMMDD-<topic>-from-x/

Confirm Collection Scope

When collecting a set of posts by search, confirm the following 4 points.

  1. Target query (hashtag / keyword / OR / domain) or known post URL
  2. Time window
  3. Target count
  4. Destination note name

If the user omits them, propose default values and confirm. Do not run arbitrarily with 4時間 / 500件 . If known post URLs are already listed, that set of URLs defines the target and count, so no time window is needed. If the destination can be uniquely determined from existing deliverables, do not ask again. For a new model name on the day of announcement, etc., the feed=latest 500 items can be exhausted in a few hours, so if the count limit is reached first, record the actual target period, report it as unfulfilled, and propose either additionally fetching the official post immediately after the announcement as a known URL, or increasing the count.

Collection Paths

  • For collecting public posts, use GET https://api.fxtwitter.com/2/search?q=<URL-encoded-query>&feed=latest&count=100 as the default only within the agreed time window and count.q may URL-encode not only hashtags but also ordinary keywords,OR, and domains. Pass the next page to cursor.bottom を cursor . Refetch known posts with GET /2/status/{id} .
  • FxTwitter is not X's official API. Do not pass cookies, tokens, or account information, and limit it to one-off collection of public posts.
  • Not just the HTTP status, but also the JSON code and adopt only200 の results .400、404、500 : do not retry repeatedly; return to the X screen or the primary source.
  • ORFor compound queries mixing punctuation, parentheses, and quotes, , posts unrelated to the topic that match only one word code: 200 Sometimes returned as such. Split into one query per phrase, search multiple times, and locally check whether the results contain the target term.
  • For reactions to an announcement, searching the official page title in quotes picks up posts containing the URL (measured: 16 title-phrase hits versus 1 hit for a generic keyword). Record like counts along with the date checked.
  • Do not use for private, deleted, or access-restricted posts, or for continuous monitoring.
  • If the API fails, or if you need to confirm the on-screen context of a post, open X live search with a browser-based tool, obtain the post URL, time, body text, and image URLs, and scroll down.
  • Even when structured extraction is weak, keep the order fixed: save raw posts, coarse classification by visible domain / card title, then deep dives into high-signal primary sources.
  • Don't insist on full automation; prioritize official accounts, live blog entry points, and representative images.

The core of this skill is not browser operation tips, but the order in which SNS noise is converged into primary information.

Workflow

  1. Fix the collection targets Decide the target query, time window, target count, main objective, and where to save the research.

  2. Collect efficiently from the API Use the API's results to perform structured extraction keyed primarily on status URLs, and deduplicate. Only supplement from the screen if the API fails.

  3. Save raw artifacts first Keep tmp/<slug>-posts-raw.json and intermediate JSON per batch.

  4. Do local classification first Create theme, domain, repo, and image candidates from rawText, tweetText, display text, and the visible domain of X cards. Sort popular candidates with likes、reposts、views after deduplication, but do not use reaction counts as grounds for accuracy.

  5. Deep-dive only into high-signal links Do not resolve every t.co; follow only high-frequency links, official accounts, high-signal posts with images, and posts where the repo name is half readable. Check implementations and claims in the repo / Docs / app / articles at external link destinations, and do not count a README's roadmap or planned items as implemented.

  6. Write research notes source-centric Rather than impressions of posts, make it the source of truth which primary information a post converged to. Divide evidence levels as A = primary artifact + measurement, B = primary artifact, C = in-post demo/self-report, D = idea only. For claims whose only basis is a video UI, attribute them as動画内の説明ではand do not treat implementation, accuracy, tokens, or cost as confirmed.

  7. Store images in two layers. As a guideline, keep the curated set to 5–10 images and bulk to about 20–30. For X Articles, check the title / blocks / cited sources, for images check the original resolution, and for videos check the captions or a representative frame, and do not assert the content of media you have not viewed. When posting in an article, prioritize the minimal frames needed to support the claim, such as before / operation / after, over redistributing the entire piece, and remove unrelated browser chrome. Do not adopt card images that contain outdated star counts or pre-implementation descriptions unless they have value beyond a link to the primary artifact. If an image alone would be misleading, add a note making clear it was added by the author, or do not include it.

  8. Finish with reproducible artifacts. Prepare research notes, raw JSON, key primary-source URLs, image save locations, and if necessary append to the manifest. If external media has been saved or processed, record the original post URL, retrieval date, source-relative path, and SHA-256 in the manifest, and verify them with builder / audit.

Branching Rules

  • Public, one-time, bounded collection: use the FxTwitter API search,code: 200 and keep only the results in the raw artifact.
  • If the API fails or the post's on-screen context is needed: prioritize official accounts, live blogs, official blogs, and posts with a repo link.
  • If resolving shortened URLs is heavy: do not resolve all of them. Resolve only high-frequency, high-signal ones.
  • If there are too many images: secure the curated ones first, and cap the bulk.

Quality Gates

  • Both the collection count and the time window are satisfied. If they cannot be satisfied, the actual target period and the uncollected range are clearly stated in the research notes.
  • The main themes are organized as a body of public information.
  • If a repo name is truncated, it is corrected using the body text, card title, or official blog.
  • Image save locations are organized under research.
  • No assertions have been made that alter the original public posts.
  • When FxTwitter is used, the sources are not the API URLs but each status.url 's original X post URL.
  • Popularity ranking and evidence level are separated, and self-reported claims and ideas are not treated as proven.
  • Check the current state of the repo / app, and distinguish implemented / roadmap / author benchmark / author's own measurement. For the counts within the repo cited by posts or articles, roughly verify using gh api search/code の total_count etc., and clearly state that they are estimates.
  • The dates and claims of the posted images match, and author-added notes are not conflated with the original image's expression.

Efficiency Rules

  • tmp/*.json Save intermediate results to avoid re-fetching.
  • Do local classification first, then external fetch after.
  • If there is a cross-cutting article on the official blog, use it as a hub and descend to individual articles.
  • Do not take on too many browser-specific tips. X browser operations and general-purpose browser automation are outside the scope of this skill.

Example Prompts

  • /x-post-research "GPT-6 Luna" OR "GPT6 Luna" を直近7日で500件集めて、記事に足せる第三者評価や注意点を拾って
  • /x-post-research #MSBuild と #MicrosoftBuild を直近4時間で500件以上集めて、一次情報と repo と画像を research に保存して
  • /x-post-research #Ignite と #MicrosoftIgnite を2時間で300件、画像は代表8枚だけでいい

Related Customizations To Create Next

  • A prompt to insert X research results into keynote research.
  • A skill to select only article-suitable ones from saved images.
  • A prompt to secondarily organize only official blog / Learn / GitHub repo.