agentleFS
fluxnative-appfluxnative-app/fluxnative-claude/skills/mobile-deslop/SKILL.md

mobile-deslop

Skill0 starsChanged 10 days agoMITView on GitHub

Mobile app deslop for React Native and Expo. Changes the code to remove the generic AI-made look (purple gradients, card soup, tracked-capital labels, emoji icons, letter discs, dot-joined meta lines, competing buttons, spinner-only loading, dead buttons, hard-coded colours, radii and font sizes). Counts tells in the code, checks light and dark screenshots, names each tell by ID, fixes the worst tier first without changing what the app does, re-renders, and reports which platform ran it and what it could not verify. Use when an RN or Expo app looks AI-made, generic, templated, cheap or vibe-coded, or an AI tool built its UI, and the user wants it fixed, polished, cleaned up or made to look designed; when asked to remove gradients, emoji icons or card soup; or to count and fix hard-coded colours and radii (with expo-design-system, if installed). Not for websites, landing pages or web CSS, a critique or verdict with no edits, building new screens, performance, TypeScript errors or code review.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

12 rules. About 0 are common, 5 are distinctive, and 7 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 7

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Keep features, data flow and navigation unchanged: no new screens, no removed features, no changed routes.

    line 17
  2. Distinctive:

    Check before changing anything:

    line 33
  3. Distinctive:

    Before judging, write your best source at the top of the report: 1.

    line 44
  4. Very specific:

    Write which platform ran the app on Ran on:.

    line 46
  5. Very specific:

    Run the commands in references/code-counts.md and fill its count table.

    line 50
  6. Very specific:

    Follow references/screenshot-pass.md.

    line 55
  7. Distinctive:

    Never invent a tell to fill a list.

    line 64
  8. Distinctive:

    Never fix a lower tier while a higher one is open.

    line 83
  9. Very specific:

    Default tier per tell ID: the table at the top of references/tells.md.

    line 93
  10. Very specific:

    When the cause is shared (a theme value, a row component, a card wrapper), fix it there once, not per screen.

    line 96
  11. Very specific:

    Run the project's type check, lint and tests (npx tsc --noEmit, npm run lint, npm test) before the new render.

    line 100
  12. Very specific:

    Check each new render against the Hard rules in references/fixes.md (contrast: pair and ratio).

    line 103

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Mobile deslop

sniff → count (facts) → findings → fix the top tier → new render, rule each finding fixed or open → stop at 3 rounds → report, ending with a Not verified line

Scope

A tell is a choice the model made by default, not for this product. This skill finds tells and removes them. Keep features, data flow and navigation unchanged: no new screens, no removed features, no changed routes.

References

  • code-counts: load at Step 1.
  • screenshot-pass: load at Step 2; its Emit at Step 4.
  • tells: load at Step 3 to look up tell IDs.
  • fixes: load at Step 5.
  • hand-off: load for sample content, a second opinion or a new screen the user asked for.
  • verify-without-device: load when you name your evidence.
  • fluxnative-apps: load only if the FluxNative marker matches.
  • pass-log: load at Step 5 for a commit per tier.
  • fluxnative-mcp-calls: load before Step 2 only when working through the FluxNative MCP (preview_view connected: shots from its preview, or no local checkout).

Sniff

Check before changing anything:

  • package.json lists react-native or expo. If neither, stop.
  • The styling system and its token file: src/theme/, theme.ts, a tokens file, a Tamagui config, a NativeWind or Tailwind config, a shared styles module.
  • A design brief: DESIGN.md at the root. In a FluxNative app, also .fluxnative/memory/design-dna.md.
  • The FluxNative marker (last section of this file).

Rule: fix inside the existing system first. Never add a UI kit, NativeWind or a second theme. Where the brief and this skill disagree, the brief wins, except on WCAG contrast and working controls; name the conflict in the report.

Name your evidence

Before judging, write your best source at the top of the report: 1. a release build on a phone; 2. a simulator or emulator; 3. Expo Web at 393×852, 2x; 4. screenshots the user gave you; 5. the code only. Everything references/verify-without-device.md marks "cannot" for your source goes on the Not verified line. Write which platform ran the app on Ran on:. Claim only what ran there: a web render proves nothing about iOS or Android, and an iOS run proves nothing about Android.

Step 1: pre-flight count

Run the commands in references/code-counts.md and fill its count table. Keep it: the report shows it before and after. Its two screenshot counts per viewport (filled primaries, non-neutral hues) come from Step 2's baseline shots, or are n/a, code only.

Step 2: look

Follow references/screenshot-pass.md. Code only: skip the shot checks, read the JSX, mark results from code.

  • Screens: the ones the user names, else the default set there; at most 6.
  • Shoot each at 2x, in light and dark: a review shot and a show shot. These are the baseline shots.
  • On each shot, run its checks (squint, grayscale, cover the titles, bottom 120 pt, 200% text, dark mode), then walk each core flow (Walk flows there).

Step 3: name tells

Name each tell with its ID from references/tells.md: C colour, T type, L layout, D data, W words, I imagery, N navigation, S states, M motion. Never invent a tell to fill a list. Do not list a rule that holds.

Not tells, do not flag:

  • a gradient or saturated colour that is the brand;
  • sample labels the brief requires;
  • native platform defaults (system tab bar, system sheet, system switch);
  • a reference's structure the user asked for.

Step 4: findings

One row per finding, in this shape:

screen | tier | pattern | region | observation | effect | fix | evidence (seen / measured / from code)

At most 8 findings per pass across all screens, in tier order, the biggest first impression first; tier 1 is never cut. The Not listed: line, field rules and an example: references/screenshot-pass.md, Emit.

Step 5: fix in tier order

Never fix a lower tier while a higher one is open.

  1. Broken: dead controls, content under bars, clipped text, wrong figures, a part of the request missing, words, palette or brand copied from a reference the user gave.
  2. Hierarchy: nothing leads, competing primaries, a small hero figure.
  3. Structure: card soup, identical tiles, a stamped hero, web chrome.
  4. Colour: one accent with one meaning, gradients, dark mode.
  5. Surfaces: radii, one way to separate, stacked borders and shadows.
  6. Density: row heights, meta stacks, dead bands.
  7. Details: icon family, tabular figures, press feedback, copy.

Default tier per tell ID: the table at the top of references/tells.md. A finding may be raised one tier when its effect says so; write the reason in the effect field. A finding you cannot fix this turn (needs a decision, an asset or a backend) is ruled open: blocked, <reason> and does not block lower tiers. When the cause is shared (a theme value, a row component, a card wrapper), fix it there once, not per screen. references/fixes.md holds the Hard rules every fix must meet, recipes per tell group, and "Thin screens" for freed space.

Step 6: re-check

  • Run the project's type check, lint and tests (npx tsc --noEmit, npm run lint, npm test) before the new render. A failure is a tier 1 finding.
  • Take a new render of each touched screen and re-run the counts you changed.
  • Rule each finding fixed, naming what the screen now shows, or open.
  • Check each new render against the Hard rules in references/fixes.md (contrast: pair and ratio).
  • A successful write or a build that compiles is not evidence.

Stop

A round is one pass of fix, type check, new render and ruling. Stop when no tier 1–3 finding is open (other than open: blocked) and no new one appears; 3 rounds have run; only details (tier 7) remain; or the user stops you.

Report

Evidence: <source from "Name your evidence">
Ran on: <Expo Web export | iOS Simulator | Android emulator | phone (model) | not run>; claims cover only this platform
Counts: <table, before → after>
Findings: <rows, each ruled fixed (what the screen shows now) or open>
Before/after: <route, scheme: baseline shot → after shot, one line on what changed, or n/a, code only>
Files changed: <paths>
Not verified: <every "cannot" for your source, in the words of verify-without-device.md, and screens not reached>

Never

  • Copy another product's pixels, brand, copy or imagery. If a screen-library MCP is connected, take patterns from it, never pixels.
  • Fake data to fill a chart.
  • Change what a screen does to make it look better.

If this is a FluxNative app

FluxNative app: the root AGENTS.md starts with # Generated app source contract and src/platform/, src/theme/ and src/features/ exist; then read references/fluxnative-apps.md before Step 1. That file decides imported apps.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.