UI Theme QA (Godot, reductive)

SkillDev tools

Use this skill to find and fix undisciplined Control-node UI in a Godot project, spacing that follows no ladder, font sizes with no modular scale, near-duplicate colours, and text that fails WCAG contrast, by measuring the live editor with the `hera-godot` CLI and snapping theme tokens to a reference corpus. It also reports (without changing) inert container wrappers and decorative nodes. Triggers: \"clean up the UI spacing/type/contrast\", \"the UI values look arbitrary\", \"tidy the Godot UI\", \"check UI contrast\", or any request to mechanically remove statistical UI-design defects without a redesign. Reductive only; edits per-node theme overrides (undoable), never copy or layout content.

Use UI Theme QA (Godot, reductive) in Claude, ChatGPT or Ahel Desktop

Free. Sign in, add UI Theme QA (Godot, reductive) and connect your AI. About a minute.

Also: Claude Code · Cursor · Codex

Then ask your AI: use the UI Theme QA (Godot, reductive) skill

Details

Instructions available. Your AI can read the instructions. Execution depends on the setup they require.

Add Ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.

UI Theme QA (Godot, reductive)Start free

What this skill tells your AI

The instructions your AI receives, as published by hashgraph-online/awesome-codex-plugins in plugins/NotNull92/hera-agent-godot/skills/ui-theme-qa/SKILL.md and read by Ahel’s review.

Mechanically remove statistical design defects from a live Godot editor, over Control nodes. It does not redesign: it snaps theme tokens to a reference corpus and never touches copy, information order, or layout structure.

Six areas, split by what they may do:

areaseffect
enforcingspacing · type-scale · color · contrastset theme_override_* (undoable)
report-onlycontainers · decorationproposals in the report; no mutation

The report-only pair would have to delete or re-parent nodes, which cannot be decided mechanically — a node that looks decorative may be a divider, and flattening a wrapper breaks the $Path/To/Node references scripts rely on. They measure and surface; a human decides.

Why a pipeline, not one pass: one context holding every rule while fixing a large UI lets loud items crowd out quiet rules, which get silently dropped. So inspection and enforcement are split per area, and findings live in files. The orchestrator (this context) stays thin — it holds routing only; the detailed rules and findings live in the reference doc and the findings files.

Rules SSOT: references/ui-theme-areas.md. Values: references/reference-corpus.md. Read only the area section you are working; do not load all rules into one context.

0. Prep

  1. hera-godot status — confirm one live editor (the live-editor skill covers setup). If multiple, pass hera-godot --instance <pid> to every mutation.
  2. hera-godot guidance ui — respect the returned UI mode.
  3. Ensure git is clean or on a branch; each area fix is one commit.
  4. Enumerate the target Controls: hera-godot --ids scene tree (edited scene) — this is the static surface the checks read.

1. Inspect (parallel, static, read-only)

Dispatch one inspector per area (spacing, type-scale, color, contrast, containers, decoration). Each:

  • reads only its area section of references/ui-theme-areas.md,
  • measures the live editor with the read commands listed there (node get --props, eval get_theme_color, game ui tree rects),
  • writes findings-<area>.md using the schema (each item a check predicate, never a status word).

No mutation here. Inspectors snap replacement values to the corpus so the fix is carried in the finding, not invented later.

2. Report

Merge the findings-*.md into one local HTML report and serve it: python3 -m http.server <port> from the run folder, then hand the user http://localhost:<port>/. Show measured values verbatim (distinct spacing values, font-size ratios, contrast ratios). No artifact, no browser auto-open — link only.

3. Enforce (sequential: spacing → type-scale → color → contrast)

One area enforcer at a time, in order. containers and decoration are skipped here — they enforce nothing. Each enforcing area:

  • loads findings-<area>.md + its area section,
  • re-measures each check from the live editor; applies the fix only where the predicate is currently false (never trust the finding text),
  • enforces with hera-godot node set <path> --prop "<theme_override…>" --value <v> (undoable), snapping to the corpus rung/hex,
  • commits that area before the next runs.

Order = dependency order: spacing commits before type-scale, color converges before contrast, and contrast runs last because it depends on the final colours.

4. Re-inspect (fresh, parallel)

Dispatch fresh inspectors (not the enforcers) to recompute each enforcing area's check predicates from source and write verify-<area>.md (true = applied, false = missed). Any false re-enters step 3 for that area only. The report-only areas have no check — they changed nothing, so there is nothing to re-verify, and inventing a predicate that always passes would be a lie.

5. Render QA (confirmation only)

hera-godot run --current --wait, then hera-godot screenshot --runtime --analyze for a before/after and a clipping/blank sanity gate (possible_clipping, nonblank). hera-godot stop --wait. Render is confirmation, not measurement — the design facts came from steps 1/3. Update the report in place.

Contract (why this skill exists)

  • Thin orchestrator. This context holds routing only. Rules and findings live in files. Holding all areas here reintroduces the capacity drop.
  • A check is a predicate, not a status. A finding is a check recomputed from the live editor every time — never a stored "done". Enforcers fix only false predicates; fresh re-inspectors recompute them.
  • Snap, don't invent. Replacement values come from the corpus, which is rooted in Godot's own default theme plus WCAG. A project's own theme/tokens win over the corpus if present.
  • Undoable + reductive. Only per-node theme overrides change, via node set (Ctrl+Z survives). Copy, order, and layout are inviolable — which is precisely why containers and decoration report instead of enforce. Widening that would need an explicit opt-in flag, never a quiet default.
  • Static inspect, render once. Measurement is node get / eval / game ui tree; the screenshot analyzer is the confirmation gate, not a measurement tool.

Non-interactive

Fast, mechanical, non-interactive. Ambiguous items are proposed in the report, not asked. Redesign-scale decisions (which accent, is this density right) are out of scope — surface them, do not silently pick.

Companion primitives and boundaries

Project-wide Theme construction is available through resource create Theme <res://theme.tres> plus theme set, and screenshot diff can locate before/after pixel changes. They complement this skill but do not change its safe default: enforcement remains undoable per-node theme overrides, while project-wide resource writes are persistent and not undoable.

Automatic wholesale restyling remains out of scope. Hera can apply a style an agent or user has already chosen, but this workflow does not invent taste or silently replace the project's visual language. See the repo's docs/UI_THEME_QA_DESIGN.md §11.1.

Signals

GitHub stars
1k
Forks
316
Last commit
Oct 2026
Advanced
Item type
skill
Key
ui-theme-qa
Source
github.com/hashgraph-online/awesome-codex-plugins