Bilingual Pages
This skill should be used when the user asks to "add a Chinese version", "make the page bilingual", "增加中文版", or needs to maintain translated variants of generated HTML/docs without them silently drifting apart.
Design Regression Guard
This skill should be used after fixing any visual, layout, accessibility, or content defect on a website — turn each fixed defect class into an automated assertion so it cannot silently return. Use when the user says "make sure this doesn't break again", after a design-review round, or when building a check suite for a static site.
In Page Model Probe
This skill should be used when a web page embeds a numerical or physical model in client-side JavaScript (a simulation, an interactive figure, a parameter-driven visualization) and its behavior needs to be verified or tuned — sweep parameters, check monotonicity, and reconcile model vs render without touching the UI.
Results Site
This skill should be used when the user asks to "make a results website", "publish results to a page/HF Space", "展示方法和效果", or wants research results presented online concisely but with every claim traceable to evidence.
Evidence Driven Optimization
This skill should be used when the user asks to "optimize the method", "improve accuracy", "why is the metric low", or wants to iterate on an algorithm/pipeline against a measurable target without cargo-culting changes.
Single Source Visual Consistency
This skill should be used when one physical or statistical quantity appears in multiple views of the same page or figure — a schematic, a zoomed inset, a readout, a simulated sensor image — and the user reports the views "don't match" or "look inconsistent". Establishes one source-of-truth function per quantity and a measured consistency chain.
Render Verify
This skill should be used when the user generates HTML pages, figures, or visual artifacts and needs to verify they actually render correctly — "check the page", "does it look right", or after any site/figure build before publishing.
Verification That Can Fail
This skill should be used whenever you are about to trust a green check — a passing test suite, a regression harness, a comparison script, an eval — and especially before claiming work is done. Use it when the user asks "is this actually working?", "validate our results", or when a verification passes on the first try.
Measured Design Audit
This skill should be used when the user asks to review, improve, or "beautify" a website's design, layout, colors, fonts, or accessibility. Replaces eyeballing with measurement — computed contrast, touch targets, font-size census, multi-viewport overflow — so every fix is provable and every claim has a number.
Gt Validation
This skill should be used when the user asks to "validate a method/estimator", "compare with ground truth", "find datasets to validate on", or needs to evaluate a perception/estimation pipeline against labeled or proxy ground truth without fooling themselves.
Debug Ledger
This skill should be used during any multi-step technical investigation or long autonomous work session — keeps a found→fix ledger and a plan file so that failures, rejected ideas, and negative results survive context loss and become publishable evidence.
Dataset Release Integrity
This skill should be used when publishing or maintaining a multi-config dataset release (HuggingFace, Zenodo, an internal data lake) — when the user asks to "add a subset", "fix the dataset schema", "update the dataset card", "unify the configs", or whenever a README/datacard makes a structural claim that nobody checks against the data.
Release Parity Refactor
This skill should be used when refactoring, rewriting, or reorganizing code that already produced a published artifact — a released dataset, a shipped model, a public export. Use it when the user says "clean up this pipeline", "extract these scripts into a package", "modernize the build", or whenever the old code's output is already in someone else's hands.