| name | social-platform-reader |
| description | Use when Codex needs to inspect Douyin, Bilibili, WeChat public-article pages, boards, favorites, collections, posts, videos, or similar platform artifacts, and must extract only directly visible evidence from a logged-in browser or a stronger installed backend. |
Social Platform Reader
Use this skill to read social-platform content through the strongest available backend while preserving evidence quality.
Prefer installed upstream MCPs over custom parsing when an upstream backend is mature enough for the target platform and task.
For research reporting tasks, keep extraction structured enough that the result can be handed to research-docx-export without another manual cleanup pass.
Backend Order
Choose the backend in this order:
-
Use native Codex computer-use/browser capability only when it is actually exposed in the current session. Do not assume computer_use exists on every platform.
-
Use chrome-devtools for login-gated or browser-exact tasks when the MCP connection is healthy, especially boards, favorites, collections, user-visible tabs, lazy-loaded pages, WeChat article pages behind verification, or any case where the user asks to read what is currently visible in Chrome.
-
Add reusable agent-browser templates when the task should be turned into a repeatable browser-evidence workflow with saved artifacts or a reusable evidence bundle.
-
Add direct agent-browser only when the task is template development, capture debugging, nonstandard interaction, or another case that should bypass the template workflow on purpose.
-
Fall back to browser reading for Douyin, Bilibili, WeChat article pages, and similar platform tasks by default. Do not invent or require a dedicated platform backend in VELA.
Read references/upstream-backends.md when the task requires backend-selection rationale or platform extension.
Read references/agent-browser-templates.md when the task should switch from one-off reading to reusable browser automation.
Platform Routing
Douyin
Use chrome-devtools by default.
Add reusable agent-browser templates when video-page capture should be standardized, repeatable, or saved as a reusable evidence package.
Add direct agent-browser when video-page capture needs backend debugging, custom scripts, or interaction logic that the template workflow does not yet cover.
Treat captions, author names, timestamps, and visible engagement counters as browser-observed fields. Do not infer transcript text unless the page visibly renders it or the media is explicitly transcribed in-page.
Bilibili
Use chrome-devtools by default.
Add reusable agent-browser templates when a stable repeatable capture flow is more important than one-off manual inspection.
Add direct agent-browser only for debug or nonstandard interaction.
Read video title, uploader, publish time, description, tags, and visibly rendered transcript or summary only when those fields are on the page.
WeChat Public Articles
Use chrome-devtools by default.
Add reusable agent-browser templates when the article should be captured as a repeatable browser workflow with screenshots and structured artifacts.
Add direct agent-browser only for debug or nonstandard interaction.
Treat title, public account name, publish date, body paragraphs, subheadings, and source link as browser-observed fields only when they are visibly rendered on the page.
If the page shows an environment exception, verification prompt, or anti-bot interstitial, stop and report the blocker rather than summarizing the article from guesses or secondary mirrors.
Workflow
1. Identify the Artifact
Classify the user target before extraction:
-
platform: Douyin / Bilibili / WeChat article / other
-
artifact type: board, collection, favorites, profile, search result, note, post, video, public article
-
access mode: public URL, logged-in browser page, or already-open tab
If the target is ambiguous, resolve the page first and only then extract.
2. Choose the Backend
Use the routing rules above.
If a platform-specific backend is unavailable or excluded from the public VELA package, fall back to chrome-devtools when the browser can still expose the page directly.
If the task requires repeatable browser automation, prefer reusable agent-browser templates first, and switch to direct agent-browser only when the template workflow is insufficient.
If all applicable backends are unavailable, stop and report the missing prerequisite instead of improvising.
3. Use agent-browser Templates When Needed
When reusable agent-browser templates or direct agent-browser are selected, prefer the reusable template script over ad hoc commands:
Default template behavior:
-
open the target URL with platform-safe defaults
-
wait for page hydration
-
collect title and final URL
-
save an interactive snapshot JSON
-
save a visible screenshot
-
optionally save session state
-
write a metadata JSON artifact
Do not treat the template output as hidden truth.
It is still only a browser-observed artifact bundle.
The difference is only orchestration:
4. Extract Only Direct Evidence
For every item, extract only fields that are directly exposed by the chosen backend or the rendered page:
-
title
-
author
-
published_at
-
body_or_caption
-
tags_or_topics
-
media_type
-
source_link
If a field is absent, write NOT_DISPLAYED.
Do not infer hidden posts from counters, collection totals, or partial metadata.
5. Normalize the Output
Use the schema in references/output-schema.md.
Default output has two layers:
-
A summary table:
index | platform | title | author | time | 100-200 char summary | link
-
A per-item evidence block:
key points + evidence fields + short explanation
When the user asks for comment analysis, add:
-
A ranked visible-comment block:
comment rank | note title | author | like count | content | scope note
If the user explicitly wants Chinese headings, translate only the final output labels, not the extracted evidence.
6. Handle Dynamic Pages Carefully
When the page is lazy-loaded or client-rendered:
-
snapshot the current page
-
inspect visible DOM and runtime state
-
inspect relevant network requests when needed
-
ask the user to keep the page open or scroll only if new items must be rendered
If the same interaction must be repeated across many visible items, promote the task to reusable agent-browser templates first, and only drop to direct agent-browser when the template workflow no longer covers the interaction cleanly.
Report Handoff
If the user wants a deliverable file rather than only a chat summary:
-
keep the extraction structured
-
state the visible-evidence boundary
-
rank comments only from rendered visible comments
-
hand the structured result to research-docx-export
If agent-browser was used, keep the artifact directory path in the report notes so the screenshots and snapshot JSON can be audited later.
If reusable agent-browser templates were used, keep the same artifact directory path because the template workflow is only a standardized wrapper around the same browser-evidence workflow.
Failure Rules
Stop and explain the blocker when:
-
the required MCP server is unavailable
-
Chrome remote debugging is not enabled when chrome-devtools is required
-
the user is not logged in for a login-gated page
-
the page structure changed and no direct evidence can be recovered
-
the platform blocks rendering or the backend returns incomplete data
-
a WeChat article requires verification or shows an environment exception page
When blocked, report four things:
-
what page or tool was checked
-
what direct evidence was found
-
why the evidence is insufficient
-
the smallest next action required from the user
When comments are requested, also report whether the result covers:
-
only visible top-level comments
-
visible replies that were explicitly expanded
-
or the full rendered comment set after scrolling
Scripts
Use these scripts instead of rewriting setup steps by hand:
References
