This Skill records the article-type page practices decided by the user (2026-10-11) and serves only effect.video-type sites. It does not repeat the content of other Skills; it only specifies “who does what, in what order, and what is not allowed,” with details pointing to the corresponding Skill to read.
Images (cover images) are as important as the body text and are part of writing an article: every article must have a poster-style cover, using /imagegen, see references/cover.md。
Why do it this way
The value of an article lies in credibility: it writes about things others have already made clear, and we are responsible for selecting, connecting, and organizing. Self-invented numbers, effects, or evaluations are a hard failure for EEAT as soon as they go wrong. Therefore the workflow separates “finding material” from “writing”: material is first saved into a file with sources, the writing model may use only what is in this file, and then another model checks it item by item against the file.
One iron rule (all other rules derive from it)
Do not invent. Every fact, number, quotation, and evaluation in an article must point back to a real accessible source (sources/<slug>.json one of the items in ). Our own content is only of two kinds: connective and organizing language, and on-site pricing/parameters (generated by code, not handwritten). If no source can be obtained, it is better not to write the content; record it in the report as a “gap.”
Who does what
| Stage | Executor | Description |
|---|---|---|
| Any research on the X platform (finding posts, trends, community evaluations, comparison material) | Grok（fleet-go new … --to grok） |
Posts must include the original link, author, date, and whether they contain video |
| YouTube reference videos, Reddit, official blog/docs/pricing | Claude（fleet-go new … --to claude --tier sonnet, when a browser is needed opencli） |
Official and first-hand sources take priority |
| Canvas design (UI library and the “article page standard” page) | Claude(canvas agent, executor-sonnet; reading and writing the canvas requires this app’s Artifact tool) | UI library first, then pages (see /design） |
| Architecture decisions such as components/schema/deep links/data sources | Claude Opus | Complex trade-offs |
| Applying templates, page migration, mechanical editing | Claude Sonnet | |
| Drafting | Gemini 3.8 Flash（fleet-go new … --to gemini） |
Write in all the material, only connect and organize, state things positively |
| Review after drafting | Claude Sonnet | Check facts, citations, tone, and exaggeration against sources |
| Code/SEO review | Claude Opus(read-only); Sol read-only is also acceptable | maker and checker must be different people |
| Cover image(one poster-style cover per article) | Codex image_gen（/imagegen，fleet-go new … --to gpt --make image） |
Hollywood blockbuster feel: exaggerated, strong impact, clear visual focus; no text/logo/trademark/real people; see details in references/cover.md |
| Deployment and online readback | Follow agent-fleet’s deployment template; production is read-only for readback | Do not perform real paid verification in production |
Always dispatch orders using agent-fleet ’s fleet-go new(run_in_background, do not add &), see the writing brief template in references/writer-brief.md。
Process (complete in one pass, no scheduling for later)
Every step has an artifact file, written into ~/.agent-reports/<日期>/ and updates the project relay record (rankup 's relay protocol).
- Reconnaissance: current routes/components/schema/deep links/pricing data sources (read-only, Sol or Claude).
- Canvas: on the same design canvas, create/update the "Article Page Standard" page; the UI library defines components first, then draw the four templates: model page, blog, model comparison, product comparison (
/design, publish one page at a time). - Component implementation: shared components + content schema (backward compatible) +
/generatordeep links by model parameters + X embeds + shadcn charts + parameter tables and credit chart data auto-generated from in-site pricing (frontmatter only declaresmodel/xKind, no numbers). URLs and canonical stay unchanged; JSON-LD is only added to, never modified. Drafts and the manifest'sgapsdo not enter the build output. Seereferences/components.md。 - Source collection: one
sources/<slug>.jsonper article; for format and source rules seereferences/sources.md. X goes to Grok, the rest to Claude. After collection, verify item by item that links open and dates and authors match. - Writing: Gemini 3.8 Flash writes according to
references/writer-brief.md; the inputs are the sources file, in-site parameter table data, and the article skeleton. - Review: Claude Sonnet checks against the checklist in
references/writer-brief.md; changes are written back; only after passing does it enter the repository. 6.5 Cover image: one per article, generated according toreferences/cover.mdusing/imagegen, saved to disk and converted to webp, written into frontmatter (cover and og image from the same source). - Migration and integration: write into
content/articles/**, three guide cards on the page (first screen, in-article, end of article), YouTube iframe required, X comment wall, Reddit link only, Prompt collection card (built in if available), parameter table/comparison table. - Review: Opus read-only reviews code and SEO, Sonnet reviews content; fix must-fix items.
- Deployment and read-back: after production deployment, read-only read-back (page, embeds playable, JSON-LD, canonical, side-by-side screenshots); IndexNow.
- Evaluation: see
references/review-eval.md(GeFei AI scoring, AITDK monitoring, ChatGPT anonymous Q&A); if the score does not meet the bar, return to step 5.
Required page elements (model page)
First screen (cover, title, summary, fact bar, guide card) → article introduction → YouTube reference video (widescreen iframe, fullscreen playable) → parameter/price table (inputs and outputs, duration, credits per item and per second) → Prompt collection card (when the model has a collection) → community review wall (X share card, with playable video) + Reddit hyperlink → credit chart → in-article guide card → FAQ → sources (official docs/pricing pages first, independent reviews after) → end-of-article guide card → related internal links. Guide card clicks go to /generator and carry model parameters. For comparison-type articles, replace the "parameter table" with a "us vs them" price and feature table; competitor data only cites sourced collection. For order and components, references/components.md is authoritative.
Linked Skills
agent-fleet: dispatch and interjections;rankup: relay records, SEO/GEO criteria, scripts and checklists for the evaluation stage.write: de-AI-ification check for Chinese drafts (not needed for in-site English drafts);gefeiSeries: scoring and keywords;opencli: logged-in browser data retrieval;design/canvas: component design;imagegen: illustrations.- The evaluation criteria and scripts are all in rankup; do not copy a duplicate here.
How to submit after finishing
Report: the number and types of sources for each article, gaps, a summary of changes between the Gemini draft and Sonnet review, evaluation scores, and online read-back evidence. Write "unverified" for anything not verified.