skillZs
★ LIVE SKILL TAGS ★
>>> LIVE SKILLS INDEX <<<
* OPEN SOURCE *
NO LOGIN, NO TRACKING
※ REAL INSTALL DATA ※
← back to all skills
felipebasurto/unslop-ui186 installs

unslop-ui

Cut AI tells from interface and chart copy: eyebrow labels, 01/02/03 numbering, redundant subtitles, captions that explain what is already visible, leaked design notes, decorative badges, status dots, chart clutter and legends a direct label could replace. Applies /pstack:unslop to every string it keeps. With --surface it also fixes gradients and glows, off-scale spacing, extra-heavy weights and decorative motion.

How do I install this agent skill?

npx skills add https://github.com/felipebasurto/unslop-ui --skill unslop-ui
view source ↗

Is this agent skill safe to install?

  • Gen Agent Trust Hubpass

    The skill is a utility designed to refine user interface text and decoration by removing common AI-generated patterns. It employs local Python scripts to perform analysis and load external stylistic rules via dynamic context injection. No security risks or malicious patterns were identified.

  • Socketpass

    No alerts

  • Snykpass

    Risk: LOW · No issues

What does this agent skill do?

Unslop UI

Edit the visible text and decorative chrome of a web page, app screen, chart or dashboard so it stops reading as AI-generated. A good interface is used, not read: keep the words a person needs to act, and cut the rest.

Target: $ARGUMENTS

If the target above is empty or still shows a placeholder, take it from the user's message. With none there either, use the files changed in this conversation, or ask.

Scope: copy, decoration and charts by default (U1 to U16). If the arguments include --surface, also apply U17 to U20 (gradients and glows, off-scale spacing, heavy weights, motion), using the values in references/tokens.md. Strip --surface from the target before reading paths.

Writing rules from /pstack:unslop

Apply these to every sentence you keep or write. Cite them as "unslop 13", so they don't read as U rules.

!python3 ${CLAUDE_SKILL_DIR}/scripts/load_unslop.py

If the line above still shows a command instead of rules (agents other than Claude Code), run python3 scripts/load_unslop.py from this skill's folder and apply what it prints.

Where these and the UI rules disagree, the UI rules win:

  • Labels, buttons, tabs, table headers, KPI captions and badges are fragments by design ("Last 30 days", "+12% vs last month"). The rules about sentences (28, 29) apply to full sentences only: hero copy, descriptions, empty states, errors.
  • Rule 13 does not cover a lone "—" that stands for an empty value, or an en dash in a range ("Jan–Mar").

Cut first, write little

Deleting is the default fix. Where a rule says rewrite (U3, U8, U12), the new text may state only what the source already shows: the product's own copy, the names of its features, the data bound to the screen. Never add a number, a speed, an integration or a benefit you cannot point to in the code. With nothing true to say, delete the string, or leave it and list it as a task.

UI rules

Ids are stable. Cite them as U1 to U20 in the report.

Interface copy

  • U1. Eyebrow labels. Small uppercase or letter-spaced text above a heading ("PASSPORT", "FEATURES", "HOW IT WORKS"). It also comes as a pill or chip, often with a dot ("● POLAR EXPEDITIONS"), and as a caption over a menu or dropdown list ("POPULAR DESTINATIONS"). Delete it. The heading already says what the section is.
  • U2. Decorative numbering. "01 02 03", "01 // Features", "§", "//" prefixes. Delete unless the order is real and the user follows it (steps of a checkout, a numbered install guide).
  • U3. Subtitle reflex. A subtitle under every heading, or a line above the H1. Keep one only when it states a fact the heading doesn't. "Analytics / Understand your data better" loses the second line. The same goes for a blurb under every item of a menu, dropdown or card list ("Greenland / Discover majestic ice fjords"): keep it only when the names alone can't tell the items apart (plan tiers, settings), and then only for the items that need it.
  • U4. Botsplaining. Captions, helper text and descriptions that narrate what the element visibly does ("Click the button below to get started", "This chart shows revenue over time"). Delete. Empty states, error messages and confirmations of destructive actions are not deleted, but they are edited: one or two plain sentences that say what happened and what to do next, without "Oops!" or an exclamation mark.
  • U5. Leaked design notes. Text about how the screen works or why it was designed that way, long parentheticals, "(updates in real time)", "(powered by AI)". Delete it, or move it to a tooltip if a user would genuinely look for it.
  • U6. Status strings. Facts chained with "·" or "|" ("12 items · Updated 2m ago · Synced · v2.1"). Keep the one fact the user acts on. Put others where they're needed, or drop them. A fact that claims live state ("Synced", "All systems operational") has to pass the real-state test under Decoration, or it goes.
  • U7. Labeling everything. A label that repeats the placeholder, the icon, the button text or the section heading. Pick one cue. In a form the cue that stays is the visible label: cut a placeholder that repeats it (one that shows a format, "name@company.com", can stay) and helper text that restates it or says what the field is obviously for ("We'll use this email to contact you"). A constraint the user can't guess ("At least 12 characters") stays. A search field with a magnifier icon takes a placeholder or a label, not both.
  • U8. Staccato and empty slogans. "One. Big. Thing.", "Built different.", "Ship faster. Think less." Rewrite as one sentence that names the thing and what it does, in words the page already uses, or delete.

Length: a section heading, card title or caption inside an app screen stays at 5 words or fewer. A hero heading may be one sentence.

Decoration

The test for real state, in U6, U9 and U10: the element renders from a condition or a value ({isLive && ...}, status === "online", a prop, a query result). Markup that always renders with fixed text is decoration.

  • U9. Badges and pills without real state. "LIVE", "NEW", "AI", "Beta", "Pro", "🔥 Popular" added for flavor. Keep a badge only if it reflects data the app actually tracks (a real stream is live, an item is actually new since the last visit).
  • U10. Status dots. Pulsing or colored dots (animate-pulse, small rounded-full dots) next to text. Keep only if bound to a real live status. A loading skeleton that pulses is not a status dot.
  • U11. Decorative emojis and icons. In headings, buttons, KPI labels and feature lists, including the icon in a tinted square above each feature card. Delete. An icon stays if it is the control itself (icon button with an aria-label), if it carries meaning the text does not, or if it sits in a toolbar or in a menu of actions (edit, copy, delete) where every item has one. A menu of destinations or pages whose icons are generic pictographs (a mountain for Svalbard, a compass for Greenland) is decoration: the icon says nothing the name does not. An icon in a tinted, rounded box is the same tell and goes with its box; delete the wrapper too.

Charts and dashboards

  • U12. Chart titles that name the chart type. "Bar chart of monthly revenue" becomes "Monthly revenue". When the data is fixed (a report, a static figure) the title can be the finding instead: "Revenue doubled since March". A chart fed by live data keeps the neutral title, because a finding goes stale.
  • U13. Reading instructions. Subtitles or captions that explain how to read the chart ("Each bar represents one month", "Hover for details"). Delete.
  • U14. Saying it twice. Each quantity gets one encoding. A bar, a number on the bar, a gridline at its height and an axis tick all saying the same value is three too many; a KPI with arrow plus percent plus the word "up" plus a sentence under it ("Revenue grew 12% compared to last month") is the same mistake in text. Keep the encoding the reader needs: axis and gridlines for shape, a printed value for the number that matters (often only the last point or the max). A KPI keeps its figure, one direction cue and the period it compares against ("▲ 12% vs last month").
  • U15. Chart scaffolding that competes with the data. A frame box around the plot area, dark or dense gridlines, minor ticks, thick axis borders, 3D, drop shadows and glows on marks, background fills behind the plot. Delete them, or make what must stay (a few gridlines, one baseline) a hairline in a light gray. Data on top, labels next, scaffolding faintest. Do not redesign the chart: same type, same data, same series colors.
  • U16. Legend where a direct label works. A legend box that makes the reader match colors to names, when the name could sit at the end of the line or next to the bar. Label the series directly, ideally with its latest value ("This year 1,150"), and remove the legend. Keep a legend for more than four series or when the marks are too close to label.

Surface (only with --surface)

Each rule pairs the tell with its fix. Read references/tokens.md for the values. If you can't open it, the values in these four rules are enough.

  • U17. Gradients and glows. A gradient from one hue to another on a surface, button or active item, gradient text (bg-clip-text), colored glows (shadow-[0_0_20px_...], shadow-purple-500/50), blurred color blobs behind content. Fix: a solid neutral surface, solid text color, a neutral shadow no stronger than 0 1px 2px rgb(0 0 0 / 0.05), or a hairline border. A gradient that is only a tone shift inside one neutral can stay.
  • U18. Off-scale spacing. One-off values such as mt-[13px], py-[11px], gap: 7px. Fix: the nearest step of 4, 8, 12, 16, 20, 24, 32, 40, 48, 64. Inside a control, 2, 6, 10 and 14 are also steps. Siblings that do the same job (rows of one list, cards of one row) get the same value. Keep an odd value that exactly centers something.
  • U19. Extra-heavy weights. font-extrabold, font-black, font-weight: 800 or 900. Fix: 500 or 600, and let size and color carry the hierarchy.
  • U20. Decorative motion. transition-all, hover:scale-* on cards, animate-bounce, springs and overshoot on simple UI. Fix: transition named properties (transition-colors) at 150ms ease-out, keep a press scale of 0.97 to 0.98 if the control has one. Motion that explains something stays. Pulsing dots are U10.

U17 to U20 never change the palette, the type family, the type sizes, the radii, the layout structure or the order of sections. They replace a value with a better value, they don't redesign.

For before/after pairs, read references/examples.md when unsure how far to cut.

Guardrails

  • Change visible text and decorative elements only. Do not touch logic, data, colors, spacing tokens or layout structure. The exceptions are chart scaffolding under U15, which may be thinned or lightened, and U17 to U20 when --surface is on.
  • If the project already has tokens (Tailwind theme, CSS variables, a design-system package), fix toward those and not toward the fallback values in references/tokens.md.
  • Never change a screen that has nothing to cut. See the verdict in step 3.
  • When you delete an element, delete what existed only for it: its now-empty wrapper, a wrapper left holding a single child it only grouped, the margin or gap, the import or variable nothing else uses.
  • Never delete aria-label, alt, title on icon buttons, or screen-reader-only text. If cut visible text was the only accessible name, move it into aria-label.
  • A form field keeps a visible label tied to its input (for and id, or a wrapping <label>). Never swap it for a placeholder or an aria-label.
  • Text that comes from data, props, i18n files or a CMS: edit it at its source (the translation file, the fixture) or report it, don't hard-code a replacement in the component.
  • Claims you cannot verify from the code (customer counts, uptime, ratings, testimonials, customer logos): do not delete or rewrite them. List each as a task.
  • If a flagged case earns its place (real state, real order, legal text), keep it and say why in the report.

Procedure

Copy this checklist and track it:

- [ ] 1. Count: python3 ${CLAUDE_SKILL_DIR}/scripts/count_tells.py [--surface] <target>   (${CLAUDE_SKILL_DIR} is this file's folder)
- [ ] 2. Read every visible string and decorative element; tag each problem with U1-U20 or an unslop rule number
- [ ] 3. Verdict: heavy, light or none
- [ ] 4. Edit
- [ ] 5. Count again with the same command
- [ ] 6. Review the result
- [ ] 7. Report

Count. The counter is a regex heuristic. It finds candidates and gives a before/after number. It cannot see U3, U4, U5, U7, U8 or U14, which need reading, and it sees only the common code forms of U12, U13, U15 and U16. For U11 it counts emojis and Tailwind-style icon tiles, not icons in other markup. With --surface it also counts U17 to U20; it cannot tell a hue-to-hue gradient from a tone shift inside one neutral. It flags legitimate cases too. A count above 0 after editing is fine when every remaining hit is one you kept on purpose. Never delete something legitimate to reach 0. Your reading in step 2 is the real check.

Verdict, from the count and your reading:

  • Heavy. Tells in most of the rules, or the same tell repeated across the screen. Work through every rule.
  • Light. A few tells. Fix only those. Leave everything else exactly as it is.
  • None. The counter finds nothing, or only legitimate cases, and reading finds nothing that earns a cut. Say so, edit nothing, and name at most two optional refinements. A clean screen is a result, not a failure.

Review. The counter cannot check the reading rules or any guardrail, so read your diff and the edited files once more before reporting:

  • Every subtitle, caption, helper text and label still on the screen: does it pass U3, U4, U5 and U7?
  • Every string you wrote: is each fact in it in the source, and is it inside the length limit?
  • Every deletion: is there an empty wrapper, a wrapper with one child, a leftover margin or an unused import?
  • Every input still has its visible label, and every icon button still has a name.
  • Nothing outside the scope changed: logic, data, layout, and without --surface no styling.

If you can render the screen (a dev server is running, or you have browser tools), look at it: a hole where an element was, or a row with one item left, is a leftover to fix. Fix what you find, then count again.

Report. The table, then up to three lists, and nothing else. Rows for rules the counter covers use its numbers, the others use your own count from step 2. Kept is how many of the After hits you left on purpose.

RuleBeforeAfterKept
U1 eyebrow60
U4 botsplaining50
U10 status dot311
unslop 13 em dash40

Skip a list that has nothing in it.

Rewritten. Every string you wrote, old and new, with its rule. This is the part the owner has to check. Do not list deletions, the diff shows them.

"Finance. Simplified. Finally." → "Manage your team's money" (U8)

Kept on purpose. One line each, with the reason.

Tasks. Anything you found but did not edit (text from a CMS, a claim you cannot verify, a decision that is the owner's), each with its fix:

- [ ] src/i18n/en.json:42  U3 subtitle "Understand your data better": remove or replace with a fact

Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.

<a href="https://skillzs.dev/skills/felipebasurto/unslop-ui/unslop-ui">View unslop-ui on skillZs</a>