sts TeamDe-AI Your Pages
Installing and using the human-ui-copy skill, with a real audit from this site
On this page
De-AI Your Pages
AI writes pages fast, but what comes out often carries a distinct "AI flavor": copy that reports which technology it uses, cards that are all identically neat, and numbers that disagree with each other. human-ui-copy is the team's open-source skill; install it into an AI coding agent and it flags these problems one by one before delivery. Repo: github.com/wang11453/human-ui-copy.
What it fixes
The three most common symptoms:
| Symptom | Example |
|---|---|
| Implementation notes leaking into copy | A card badge reading "local ffmpeg - no upload required"; users don't care where the code runs |
| Self-praise | "A bold experiment", "pursuing perfection" - the product applauding itself |
| Over-uniform structure | Every module the same width, the same radius, the same gradient - so tidy it reads fake |
All of these examples come from real problems found when this site was audited.
Three core principles
- Sell the user's benefit, not the implementation. "AI tools running locally in your browser, files never uploaded" → "AI tools, free with no limits".
- AI is uniform; humans play favorites. Real layouts play favorites: the important thing is big, the secondary thing is small. If everything is the same size, nothing matters.
- Guard against over-correction. The skill ships with counter-constraints: don't smash every uniform thing, and never break the product just to break it.
Installing
- Claude Code and similar agents: put the repo's
SKILL.mdinto the agent's skills directory as documented; thereferences/appendix is for humans and doesn't need to come along - Other agents: paste the
AGENTS.mdsnippet into the corresponding file of your project - No agent? Still works: use it as a manual code-review checklist and walk through it line by line
Using it
Two modes, chosen by the page's nature:
| Mode | For | Checks |
|---|---|---|
| prototype | Drafts and demos | Must items only |
| delivery | Real launches | Must + judgement, everything |
Findings come in two grades: must means fix-it-or-it-doesn't-pass (like implementation notes leaking), while judgement means weigh the trade-off (like whether breaking a set of identical cards into an asymmetric layout is worth it). The agent delivers item-by-item locations and fixes.
What's else in the repo
| File | Purpose |
|---|---|
| SKILL.md | The only file you need to install; all the agent's judgement rules live in it |
| check.py | Auto-checks the mechanical items: debug leftovers, separators, absolute paths and more, across a whole site in one command |
| references/ | The human-readable appendix: full rules, sample audit reports, and this site's real audit log |
| examples/ and tests/ | Step-by-step rework walkthroughs and regression tests |
check.py also understands page grades (game, tool, marketing...): different page types get different whitelists, so a separator inside an odds table - an industry convention - isn't reported as a violation.
These are not AI flavor
Changing anything on this whitelist would be a mistake:
- Accessibility-required annotations (like
aria-hiddenon decorative icons) - Functional technical facts (like "the model is about 7MB" - users need it to estimate the download)
- Emphasis grounded in real user intent
Case study: a live audit of this site
We ran this skill across all of stsqz.com, and the repo's references/real-audit-static-site.md records the whole process. The biggest finding: of 461 SVGs on the site, 395 decorative icons were missing aria-hidden. That one sits right on the whitelist's edge - it's a legitimate accessibility fix worth doing, but changing all 461 in one sweep was too risky, so we batched by page, fixing and verifying one batch at a time. The rest of the repairs were exactly the table above: implementation notes and self-praise swept out.
Get involved
The repo is open source under MIT: github.com/wang11453/human-ui-copy. Found a false positive, or want to add a judgement rule? Issues and PRs are welcome.