- name
SKILL.md
Beceri dosyası yayıncının deposunda kalır.
Kopyalanan komutu bir terminale yapıştırın. Bu SKILL.md dosyasını ilgili ajana kurar.
npx skills add AThevon/genjutsu --skill paint -a cursor -yKopyalanan komutu bir terminale yapıştırın. Bu SKILL.md dosyasını ilgili ajana kurar.
npx skills add AThevon/genjutsu --skill paint -a claude-code -yKopyalanan komutu bir terminale yapıştırın. Bu SKILL.md dosyasını ilgili ajana kurar.
npx skills add AThevon/genjutsu --skill paint -a codex -yKopyalanan komutu bir terminale yapıştırın. Bu SKILL.md dosyasını ilgili ajana kurar.
npx skills add AThevon/genjutsu --skill paint -a grok -yKopyalanan mesajı bir Grok Bot sohbetinde gönderin. Bot bu SKILL.md dosyasını GitHub'dan alır ve tüm Bot'larınızın paylaştığı becerilere ekler.
Install this skill: https://github.com/AThevon/genjutsu/tree/HEAD/skills/paintYayıncıdan
Paint a complete visual universe. Brainstorm first, design system second, implement third, audit last. This is NOT a quick beautifier - it's a full design pipeline.
This skill speaks in two registers:
During execution - light ninja flair, signature, immersive. Short.
In reports / final summaries / audit results - plain, factual, dev-readable. Drop the flair entirely.
The flair lives at the intro and during work narration. The moment a result lands or a question gets asked, it's gone.
/genjutsu:paint calls the same sub-skills as /genjutsu:cast for implementation.
A whole site or web app built from real material is the job of the third pipeline,
/genjutsu:bunshin: a team of subagents under one art director, with independent reviews
until only minor issues remain or the tier's round cap is reached. paint proposes it right after
the stack scan when the brief is that size, the stack is web or there is no project yet, and the
host can spawn subagents, and switches only on a yes.
framer-motion to motion uninvited.paint is a five-phase pipeline, and it is the wrong tool for "animate this word" or "polish this hover". Those belong to /genjutsu:cast, which is the default entry point.
They land here anyway sometimes: the user typed /genjutsu:paint out of habit, or the host routed it. Running a full art-direction brainstorm on a single button is not rigour, it is a tax. Recognise the case and shorten, out loud.
It is light scope when all three hold:
If two or more fail, it is not light scope. Run the full pipeline and say in one line why.
What changes:
Announce it once, so the user knows which pipeline they got and can overrule it:
"This is a single component, so I am running paint light: one question, no design system file. Say so if you want the full pipeline."
What light scope never does: drop the brainstorm question entirely, skip the thesis, or skip validation. Every gate stays. Only their number goes down.
Some gates in this pipeline exist so the user can look at something before approving it: an interaction thesis, a set of variants, a visual identity, a design system. Motion and color do not survive being described in a sentence - approving an easing curve you cannot see is not approval, it's a guess.
So before the first gate of that kind, ask how they want to see it. Then never ask again.
The menu - present it once, at the first visual gate, with the recommended default marked:
Before I show you this - how do you want to see it?
A. Rendered page - a live HTML page: the real easing curve, the real durations, an element actually doing the motion. B. Live preview - a throwaway route in your project, real stack, real tokens. Native: a
@Preview/#Previewscratch file. C. Inline - written out here in the conversation.
Recommended default - state it in the menu, never apply it silently:
The choice sticks for the whole session. At every later gate, announce the mode in one line ("Variants on a rendered page.") and go. Do not reopen the menu. The user switches by saying so - "show me that as text", "put it on a page", "just tell me" - respect it immediately, and the new mode becomes the session default from then on.
Which host is this? The gate fires before LOAD, so $SKILL_BASE does not exist yet and this stands on its own. Detect once, cheaply, then map:
if [ -d /mnt/skills/plugins ] || [ -d /mnt/skills/user ]; then
GENJUTSU_HOST=claude-ai
elif [ -d /mnt/.claude/skills ] \
|| [ -n "$(find /sessions -maxdepth 6 -type d -path '*/.claude/skills' 2>/dev/null | head -1)" ]; then
GENJUTSU_HOST=cowork
elif [ -n "${CLAUDE_PLUGIN_ROOT:-}" ] || [ -d "$HOME/.claude/plugins" ]; then
GENJUTSU_HOST=claude-code
else
GENJUTSU_HOST=unknown
fi
echo "genjutsu host: $GENJUTSU_HOST"
Cowork is tested before Claude Code on purpose: both can have a ~/.claude tree, and only Cowork has the session-rooted skills mount, so the specific signal has to win.
Producing the preview - resolve the host, degrade, never fail:
A, by capability. Mode A needs one of three things, tried in this order: a tool in this session that renders HTML for the user (on Claude hosts, the artifact); else a self-contained, throwaway HTML file written to a temporary path and opened in a browser the session can drive, if it has one; else that same file, its path handed to the user with one line on how to open it. Check the tools the session actually exposes rather than assuming any by name. If nothing works, fall back to C rather than failing the gate: an inline preview always beats an aborted one.
B - live preview needs a project to write into. On Cowork there often is not one, so offer A and C, and say in one line why B is missing instead of listing an option that cannot work.
What goes in it. A preview that restates the sentence in a nicer font is worthless. Carry what a sentence cannot:
Rules the preview obeys:
Interaction thesis: or Visual thesis:), then its
Allowed patterns: line, says in one line that the page is the proposal and not the build, and
ends with the validation question.
It holds no implementation: code starts in a later turn, after a yes. Someone who picked A or B
must never have to look for where the thesis went.not rendered in this session. Never swap it for a face the session can render: what the sandbox can fetch is not what the site will ship.Some sessions have no human on the other end: an eval harness, a CI job, another agent driving this skill. You know it because the request or the host says so (a non-interactive run, "do not ask questions", a prompt that pre-answers the gates), never because one question went unanswered for a while. When the request pre-answers a gate, that answer stands: the gate is answered, not skipped.
In such a session every gate still produces its output. What changes is that nobody validates it:
Everything else holds: the thesis is written before any code, the modules are loaded, and the audit reports evidence. A headless run skips the waiting, never the work.
Tell the block where this skill lives. The block's first line reads GENJUTSU_SKILL_DIR.
Claude Code fills it in by itself. On any other host, put one line in front of the block, in the
same shell call, naming the directory you read this file from:
GENJUTSU_SKILL_DIR='/absolute/path/to/paint'. When the genjutsu router printed a
GENJUTSU_SKILL_DIR= line, use that value as it is.
This block defines shell state, and shell state does not survive between shell calls.
$SKILL_BASE, load_skill and load_ref exist only inside the single shell call that ran
this block. Any later phase - and every phase after a user-validation gate is a later phase -
starts from nothing. So: re-emit this whole block in the same shell call as the
load_skill lines you are about to run. Never cat "$SKILL_BASE/..." in a call that did
not define it: the path resolves to /<name>/..., the cat fails, and the module is lost.
If the block prints could not find the genjutsu modules, stop the pipeline. Show the
user the message, install command included, and do nothing else: every later step depends
on the modules, and running without them produces exactly the generic output this skill
exists to prevent.
GENJUTSU_SKILL_DIR="${GENJUTSU_SKILL_DIR:-${CLAUDE_SKILL_DIR}}"
# Resolution order, first hit wins:
# 0. claude.ai, /mnt/skills/plugins (or the older /mnt/skills/user): the
# genjutsu bundle.
# 1. This skill's own directory: its _jutsu, or the _jutsu next to it.
# 2. ${CLAUDE_PLUGIN_ROOT}/skills/_jutsu, when Claude Code substituted it.
# 3. Bounded probes: .claude/skills and .agents/skills from $PWD upward, then
# the skills directories installers write to, then /sessions (Cowork).
# 4. The Claude Code plugin cache, newest version. Last on purpose: an old
# plugin install must never win over a newer bundle.
# Resolved from scratch every time. A cache was tried here and removed: after a
# plugin update the old version directory is still on disk, so a cached path
# passes an "is it a directory" check and serves the previous release.
# No function below names a positional parameter (a dollar sign followed by a
# digit, or the whole argument list). Claude Code replaces those tokens in a
# skill's text with the words typed after the slash command, before the model
# reads this block: after "/genjutsu:cast make the cards feel physical", a path
# built from a function's first parameter would read "the/motion-principles/...",
# the second word typed. Each function reads its arguments with a bare
# `for name; do`, which walks the positional parameters without writing any out.
# A _jutsu directory counts only if it holds motion-principles. The shared
# skills directory of npx serves about 80 agents, so a directory name proves
# nothing. The entry file is named SKILL or GUIDE depending on the artifact
# (the bundle renames it at packaging time), so the name is assembled from
# parts: spelled out in full, the packaging step would rewrite it too.
genjutsu_is_jutsu() { # <candidate _jutsu directory>
for jutsu_dir; do
for d in SKILL GUIDE; do
[ -f "$jutsu_dir/motion-principles/$d.md" ] && return 0
done
done
return 1
}
# Print the first genjutsu _jutsu among the candidate paths read on stdin.
genjutsu_first_jutsu() {
while read -r jutsu_c; do
genjutsu_is_jutsu "$jutsu_c" && { printf '%s\n' "$jutsu_c"; return 0; }
done
return 1
}
SKILL_BASE=""
# 0. claude.ai mounts uploaded skills under /mnt/skills/plugins/<name>/ (seen on
# 2026-09-28), next to the user's other skills; /mnt/skills/user before that.
# GENJUTSU_CLAUDE_AI_ROOT stands in for /mnt/skills in the test suite only.
# Only a _jutsu holding motion-principles counts: a mount without genjutsu in it
# must never pass for a resolved base.
genjutsu_claude_ai="${GENJUTSU_CLAUDE_AI_ROOT:-/mnt/skills}"
for claude_root in "$genjutsu_claude_ai/plugins" "$genjutsu_claude_ai/user"; do
[ -z "$SKILL_BASE" ] && [ -d "$claude_root" ] || continue
SKILL_BASE="$(find -L "$claude_root" -maxdepth 2 -type d -name _jutsu 2>/dev/null | genjutsu_first_jutsu)"
done
# 1. This skill's own directory. Empty means unknown: never probe "/_jutsu".
if [ -z "$SKILL_BASE" ] && [ -n "$GENJUTSU_SKILL_DIR" ]; then
SKILL_BASE="$(printf '%s\n' "$GENJUTSU_SKILL_DIR/_jutsu" "$GENJUTSU_SKILL_DIR/../_jutsu" | genjutsu_first_jutsu)"
fi
# 2. Claude Code plugin root.
if [ -z "$SKILL_BASE" ] && [ -n "${CLAUDE_PLUGIN_ROOT}" ]; then
SKILL_BASE="$(printf '%s\n' "${CLAUDE_PLUGIN_ROOT}/skills/_jutsu" | genjutsu_first_jutsu)"
fi
# 3. Bounded probes, following symlinks (npx links .claude/skills/<name> to
# .agents/skills/<name>). The case guard stops the walk at "/", "." or "".
probe_dir="${PWD:-$(pwd)}"
probe_n=0
while [ -z "$SKILL_BASE" ] && [ "$probe_n" -lt 24 ]; do
probe_n=$((probe_n + 1))
SKILL_BASE="$(find -L "$probe_dir/.claude/skills" "$probe_dir/.agents/skills" -maxdepth 2 -type d -name _jutsu 2>/dev/null | genjutsu_first_jutsu)"
case "$probe_dir" in /|.|"") break ;; esac
probe_dir="$(dirname "$probe_dir")"
done
for probe_root in "$HOME/.agents/skills" "$HOME/.claude/skills" "$HOME/.codex/skills" \
"$HOME/.cursor/skills" /mnt/.claude/skills; do
[ -z "$SKILL_BASE" ] && [ -d "$probe_root" ] || continue
SKILL_BASE="$(find -L "$probe_root" -maxdepth 2 -type d -name _jutsu 2>/dev/null | genjutsu_first_jutsu)"
done
if [ -z "$SKILL_BASE" ] && [ -d /sessions ]; then
SKILL_BASE="$(find -L /sessions -maxdepth 8 -type d -path '*/.claude/skills/*/_jutsu' 2>/dev/null | genjutsu_first_jutsu)"
fi
# 4. Claude Code plugin cache, newest version first. Numeric version
# directories only, so a bare marketplace clone never wins.
if [ -z "$SKILL_BASE" ] && [ -d "$HOME/.claude/plugins/cache" ]; then
SKILL_BASE="$(find "$HOME/.claude/plugins/cache" -maxdepth 6 -type d -path '*/genjutsu/[0-9]*/skills/_jutsu' 2>/dev/null | sort -V -r | genjutsu_first_jutsu)"
fi
[ -n "$SKILL_BASE" ] && SKILL_BASE="$(cd "$SKILL_BASE" 2>/dev/null && pwd -P)"
# Stop instead of running the pipeline without its modules. Name every root that
# was tried, so a new layout can be reported instead of guessed.
if [ -z "$SKILL_BASE" ]; then
echo "genjutsu: could not find the genjutsu modules (a _jutsu directory holding motion-principles)." >&2
echo " genjutsu's pipelines do not work without them. Install the full bundle:" >&2
echo " any agent npx skills add https://genjutsu.athevon.dev -g" >&2
echo " Claude Code /plugin marketplace add AThevon/genjutsu, then /plugin install genjutsu" >&2
echo " claude.ai upload genjutsu.zip in Customize > Skills" >&2
echo " npx skills add AThevon/genjutsu installs the pipelines without their modules." >&2
echo " Tried: /mnt/skills/plugins, /mnt/skills/user, GENJUTSU_SKILL_DIR (${GENJUTSU_SKILL_DIR:-empty}) and its parent," >&2
echo " \$CLAUDE_PLUGIN_ROOT, .claude/skills and .agents/skills from \$PWD upward," >&2
echo " ~/.agents/skills, ~/.claude/skills, ~/.codex/skills, ~/.cursor/skills," >&2
echo " /mnt/.claude/skills, /sessions, ~/.claude/plugins/cache." >&2
echo "genjutsu: stop the pipeline here and show this message to the user." >&2
return 1 2>/dev/null || exit 1
fi
echo "genjutsu: modules from $SKILL_BASE" >&2
# Print a module's entry file. A missing module does not stop the pipeline (a
# partial claude.ai upload is legitimate), but it is announced, and the final
# report lists it: shell state does not survive, so the model keeps the list.
load_skill() { # <module>
skill_name=""
for skill_arg; do skill_name="$skill_arg"; break; done
for d in SKILL GUIDE; do
if [ -n "$skill_name" ] && [ -f "$SKILL_BASE/$skill_name/$d.md" ]; then
cat "$SKILL_BASE/$skill_name/$d.md"
return 0
fi
done
echo "genjutsu: sub-skill '$skill_name' NOT LOADED - not found under $SKILL_BASE. Carry on, and list it under 'Modules not loaded' in the final report." >&2
return 1
}
# Print one reference file of a module: load_ref <module> <path inside it>.
load_ref() { # <module> <path inside it>
ref_module=""; ref_path=""; ref_n=0
for ref_arg; do
ref_n=$((ref_n + 1))
case "$ref_n" in
1) ref_module="$ref_arg" ;;
2) ref_path="$ref_arg"; break ;;
esac
done
if [ -n "$ref_module" ] && [ -n "$ref_path" ] && [ -f "$SKILL_BASE/$ref_module/$ref_path" ]; then
cat "$SKILL_BASE/$ref_module/$ref_path"
return 0
fi
echo "genjutsu: reference '$ref_module/$ref_path' NOT LOADED - not found under $SKILL_BASE. Carry on, and say so in the final report." >&2
return 1
}
All sub-skills are loaded via load_skill <name> (defined above), which prints the module's
entry file, or a NOT LOADED line when the module is missing, and carries on: keep the list, the
final report needs it. When the block cannot find the modules at all, it stops, and so does the
pipeline. Every phase below that loads something must re-emit the resolution block in the same
shell call: the phases are separated by user gates, and nothing carries across them.
This is the foundation. Rush it and everything downstream is wrong. The goal: understand the user's vision well enough to write two theses they'd agree with without hesitation.
Before asking the user about tech stack, scan the project to detect what's already there:
# 1. Web (existing)
cat package.json 2>/dev/null | grep -E '"(gsap|motion|framer-motion|three|@react-three/fiber|@react-three/drei|animejs|popmotion|lenis|locomotive-scroll)"'
cat package.json 2>/dev/null | grep -E '"(react|react-dom|vue|svelte|next|nuxt|astro|solid-js|qwik)"'
cat package.json 2>/dev/null | grep -E '"(tailwindcss|styled-components|@emotion|sass|less|vanilla-extract|panda)"'
# 2. Android / Compose
ls build.gradle.kts build.gradle settings.gradle.kts settings.gradle 2>/dev/null
grep -rE 'androidx\.compose|implementation\("androidx\.compose' build.gradle* settings.gradle* 2>/dev/null
# 3. Compose Multiplatform / KMP
grep -rE 'org\.jetbrains\.compose|kotlin\("multiplatform"\)|id\("org\.jetbrains\.kotlin\.multiplatform"\)' build.gradle* settings.gradle* 2>/dev/null
# 4. Apple / SwiftUI
find . -maxdepth 1 \( -name '*.xcodeproj' -o -name '*.xcworkspace' -o -name Package.swift \) 2>/dev/null
grep -lE 'import SwiftUI|@main.*App' --include="*.swift" -r . 2>/dev/null | head -1
# 5. Apple platform sub-detection (iOS vs macOS)
grep -E '\.iOS\(|\.macOS\(' Package.swift 2>/dev/null
grep -E 'SDKROOT = (iphoneos|macosx)' *.xcodeproj/project.pbxproj 2>/dev/null
# 6. Mobile web indicators
grep -rE 'viewport.*width=device-width|@media.*pointer:\s*coarse|@media.*max-width' --include='*.html' --include='*.css' --include='*.scss' . 2>/dev/null | head -3
ls public/manifest.json public/sw.js 2>/dev/null
# 7. Legacy bridge indicators (mention in DISCOVER, do not auto-load)
find . -maxdepth 1 \( -name '*.xib' -o -name '*.storyboard' \) 2>/dev/null
find . -path '*/res/layout/*.xml' 2>/dev/null | head -1
grep -rE 'setContentView\(R\.layout' --include='*.kt' --include='*.java' . 2>/dev/null | head -1
Map the results:
motion, framer-motion, three/@react-three, anime.js, or none.
motion and framer-motion are the same library at two names. Framer Motion was renamed
to Motion; motion is the current package and framer-motion is the legacy one, still widely
installed and still published. Note which of the two is in package.json - the import path
differs and the sub-skill needs to know. If both are present, the project is mid-migration:
say so and follow whichever one the file you are editing already imports..xib, .storyboard, layout XML, setContentView(R.layout.*). Mention only, no auto-load.When the job is a whole site: propose bunshin. genjutsu has a third pipeline for the one job
this one is too small for. bunshin builds a whole website or web app with a team of subagents
under one art director, with Impeccable drawing the direction when it is installed, and it loops
independent reviews and fixes until only minor issues remain, the tier's round cap is reached, or
a round fixes nothing. It is heavy, so it is proposed, never started on its own.
Propose it once, right after the scan and before anything else you would ask, when all three of these hold:
Agent or Workflow);and at least one of these:
Before writing the proposal, check whether Impeccable is installed:
for d in .claude/skills/impeccable .agents/skills/impeccable "$HOME/.claude/skills/impeccable" "$HOME/.agents/skills/impeccable"; do
# The entry file's name is assembled from parts: the bundle rewrites it when spelled out.
e="$d/SKILL"
if [ -f "$e.md" ]; then echo "impeccable: $d"; break; fi
done
The proposal is one message, then wait:
"This is a whole site<, from scratch | , a full redesign>. genjutsu has a pipeline for exactly that:
bunshin. It harvests your real material, researches with parallel agents, lets Impeccable draw the direction (<installed here | not installed: bunshin runs without it, better with it>), builds the pages with parallel agents, and reviews and refines them until only minor issues remain, the round cap of the tier you pick is reached, or a round fixes nothing. It asks you twice: once to start, with a budget, and once to choose the direction. One run has been measured, a seven-page site: about 10.5 million subagent tokens over six to eight hours. The tiers are estimated from it, from about 6 to about 13 million, and scale with the pages. Saybunshinto switch, or I carry on with this pipeline."
genjutsu:bunshin; otherwise read its entry file next to this one, in
$GENJUTSU_SKILL_DIR/../bunshin/ (the one Markdown file at the top of that folder), and follow
it from its first step.Declare your read before the first question. The scan has run and the request is in front of you. Before the first brainstorm question, say what you already believe in one line, so the user corrects a wrong premise before it steers every question after it:
"My read so far: for , heading toward , on . Correct me before I ask anything."
<what> is the product and <whom> the audience as the request states them, <direction> the
mood the request or the existing design suggests, <stack> what the scan found. A slot you cannot
fill is said as unknown ("for an audience I can't tell yet"), never guessed; the unknown slots are
where the brainstorm starts. The line is its own message: send it and wait. A correction replaces
the slot it names; a go-ahead means the read stands. It is not a brainstorm question and does not
count against light scope's single question, and it is not a thesis: nothing in it is validated
until Phase 2. When no human is in the session to answer, state the read and carry on with it as
written.
If legacy mixed detected (XIB / storyboard / layout XML / setContentView(R.layout.*)):
Ask exactly one question during brainstorm:
"I see your project mixes [XML layouts / XIBs / classic Activities] with modern UI. For this task, should I stay on pure [Compose/SwiftUI], or integrate into a legacy screen?"
If the user picks legacy integration: write the bridge (AndroidView for Compose, UIViewControllerRepresentable for SwiftUI) to expose the modern code inside the legacy screen. Never generate new legacy code (no XML, no XIB, no setContentView).
If the project already has a visual identity (CSS variables or a Tailwind theme, Theme.kt /
Color.kt, Color+App.swift or an asset catalog, a logo, a live brand):
First, take the inventory of what is already there. On a web stack, in one shell call,
re-emit the skill-base block and run python3 "$SKILL_BASE/design-audit/scripts/audit.py" . --group tells.
Show the result as one short block, What this project already does by reflex, counted by
family, with two or three file:line examples each. Each of them is settled in the theses,
according to the mode the question below sets: in preserve mode they stay by default and go into
the Allowed patterns: line as the brand's own; in partial mode they go by default inside the
areas the user named; in redesign mode they go by default everywhere. The user can keep or drop
any one of them by name at the thesis gate.
Ask exactly one mode question during brainstorm, right after the legacy question when both apply:
"This project already has a look. Should I preserve the brand and build within it, change part of it, or redesign it?"
The chosen mode goes into the visual thesis as one explicit clause ("preserving the existing palette and logo"), so the Phase 2 gate validates it with everything else. What each mode lets the thesis replace is in "Existing Project Protocol" below. Light scope does not ask: no visual identity is at stake there, so it is preserve by definition.
The five domains to cover:
How to ask: One question at a time, starting with the least obvious domain. If you already know the tech stack from scanning package.json, don't ask - start with mood or audience instead. Each answer reshapes how you ask the next question.
How to handle vague answers:
When the user says "modern" or "clean" or "I don't know, just make it nice":
Never interpret a vague answer as confirmation. "Yeah something like that" means dig deeper - ask which part of "that" resonates.
When the user pushes to skip or rush brainstorm:
Do NOT capitulate. Instead:
"We've covered [covered areas]. I'm still missing [missing areas], which will directly impact [concrete consequence]. Want me to ask one more question, or would you rather I make assumptions and you correct them afterward?"
This gives them an informed choice. If they choose assumptions, name each assumption explicitly in the thesis.
Never negotiate the number of remaining questions ("just two more, I promise"). You don't know how many you need until you hear the answers.
When to stop: When you can write both theses (visual + interaction) and you'd bet money the user will say "oui parfait." If you'd be guessing on even one aspect, keep asking.
From the brainstorm, produce two theses:
First, clear the ground, with no extra question (light scope and preserve mode skip this; in partial mode it covers only the areas the user named). Name in one line each the page or screen this category publishes by routine, and its predictable opposite: the routine with everything taken out is one too. Then list five concrete worlds the design could draw from, taken from the product's one mechanism and its user's real scene (objects, places, documents, graphic traditions, never another brand's look), in the order they come, before choosing. The first two are the ones any run would reach: set them aside with the routine and its opposite, and write both theses from one of the other three. A world, a reference or an idea the user brought or confirmed in the brainstorm is never a candidate: it is kept as given, the list covers only what they left open, and at the gate they can call back by name anything set aside. Print the outcome under the theses, after the Allowed patterns: line, and with them in the final report, as one line: Set aside: <routine>; <opposite>; <world 1>; <world 2>. Kept: <world>.
One sentence for the visual identity, then a labelled First screen: line under it, as long as its answers need. Where a gate or a preview asks for the thesis in one sentence, that is the identity sentence, with the labelled line under it. Must explicitly address all five:
display face: undecided, chosen at the design-system gate; serif/sans/mono, weight usage, size contrast. The system face is a choice like any other and carries its reason too; when no chosen face can be loaded, say what gives the type its voice without oneAllowed patterns:; an invented customer, metric or quote is a tell, kept only when Allowed patterns: names it); on a wide frame (on native, the expanded size class or landscape when the app supports them, otherwise compact only), what takes the width, what stays at reading measure, and what stands whole before any scroll; on the narrowest frame it will meet, the form it takes so that the part carrying the idea stays whole and in view, each value beside what names it; and at rest, before any input or motion, it already shows the idea. Words alone is an answer only with what then carries the screen. An example takes its values from the brief or the project. When the brief says to invent no figure, it carries no figure the brief does not give, and shows the mechanism with the brief's own material and words (the names of the things compared, the kinds of mismatch, the states), unless the thesis says otherwise with its reason. In preserve or partial mode, it describes the first screen as it stands and what this run adds to it, without recomposing itSelf-check: read your thesis back. If any of the five areas is missing or vague ("nice typography", "a clean hero"), rewrite it before presenting.
Guessability test: could someone write this thesis from the product category alone (a fintech gets navy and a clean sans, a design studio gets an editorial serif on paper)? Then it is the category's default, not a decision: rewrite it from what is true of this product. That default with everything taken out is guessable too.
A single sentence that captures the motion and interaction language, then a labelled The one motion: line under it. Must explicitly address all five (four on light scope):
none, with what states the idea in its place. Not asked on light scopeCross-platform thesis examples:
Self-check: read your thesis back. If you can't immediately derive the CSS/JS properties from it, it's too vague. Rewrite.
Under the two theses, one line naming what it allows. Allowed patterns: followed by every
deliberate device the design relies on that a reader could take for decoration or habit: mono
labels, hairline rules, a paper ground, grain, a glow, a looping animation, a code label, a live
clock. Write Allowed patterns: none when there are none. The audit reads this line: a tell
counts as allowed only when it is listed here, so a mood word never lets one through, and the
user sees the list before saying yes.
This is the first visual gate. Offer the preview menu (see "Showing Your Work" above), then present both theses in the chosen mode. The visual thesis in particular is worth far more shown than described - an accent colour is a guess until it sits next to the neutrals.
Wait for explicit user validation of BOTH theses before moving on. If the user pushes back, don't start over - ask what feels wrong and adjust.
Once both theses are validated, read three values off them. Phase 3 passes them to search.py
as 1-10 dials:
Derive each value from a clause of the validated theses, never from the raw brief and never from a default. "Generous whitespace" settles density low; "fast and dry, no scroll reveals" settles motion low. A dial the theses do not settle is left out of the call entirely: an unsent dial changes nothing, a guessed one biases the lookup toward something nobody approved. Write each value down with the clause it came from; they are shown with the design system before Phase 4.
Light scope has no Phase 3, so it sets no dial.
On a web stack, load tells now, never before the theses are written (a thesis written with
the catalogue open is made of what it left), in a shell call of its own, before the design-system
query. Light scope never reaches this phase. Re-emit the resolution block in that call:
# (skill-base block re-emitted above this line)
load_skill tells
load_ref tells references/web.md
Load the ui-ux-pro-max sub-skill and run it. Phase 2 ended in a user gate, so this is a
new shell call and $SKILL_BASE no longer exists. Re-emit the resolution block from "Sub-skills
Path Detection" above in this same call, then:
load_skill ui-ux-pro-max
# Query it with the validated visual thesis, not with the raw user request. The thesis is the
# thing that was approved; the request was not. Product type, industry and the mood adjectives
# from Phase 2, in that order, work best.
# Append only the dials Phase 2 settled (see "The three dials"); drop each flag the theses
# left open, and never send a dial with a made-up value.
python3 "$SKILL_BASE/ui-ux-pro-max/scripts/search.py" \
"<product type> <industry> <mood adjectives from the visual thesis>" \
--design-system -f markdown \
--variance <V> --motion <M> --density <D>
-f markdown is not optional. The default ascii format emits raw ANSI colour escapes that
survive the pipe and land in context as garbage, at roughly 3.3x the tokens for the same content.
What the dials do on this path. With --design-system -f markdown, --variance biases which
style the lookup picks and --motion attaches a motion snippet of the matching intensity: both
change the proposal. --density does not. On this path it only prints its label under "Design
Dials": the spacing scale it maps to is written by --persist alone, and paint does not persist.
So derive the spacing scale in MASTER.md by hand, from the spacing philosophy of the visual
thesis, and let the density label say what that scale has to feel like. The motion snippet is
GSAP code: on any other stack read it for its duration and easing only, and never install GSAP
because of it.
What comes back is a candidate palette with role names and CSS variable names, a font pairing with a ready Google Fonts URL, an effects note and a list of anti-patterns for the style. Treat it as a proposal, not an answer: it is a lookup against a static dataset and it has never seen the project. Keep what serves the validated visual thesis, discard what fights it, and say in one line what you took and what you dropped. A palette that contradicts the thesis the user approved loses to the thesis every time.
Filter it twice before keeping anything: through the thesis, and through tells, loaded just
before this call on a web stack (on Compose and SwiftUI tells is not loaded, and the thesis
filters alone). The dataset matches keywords, so it hands back exactly the reflex picks tells
describes:
tells lists as a reflex is held to the same test: named in
the thesis, or dropped. The one line that says what you took and dropped also says what this
filter removed, and what the thesis puts where it stood.A face this session cannot render is still the face. A blocked font request or a sandbox with no network is a fact about this session, not about the site. When the user has approved a web face (by name, or by saying a Google Fonts link or a hosted file is fine) or the project already loads one, never trade it for a face you can render here: that is the system face winning by default, and the thesis loses a decision it made. Never add a third-party font request the user has not agreed to, either: ask once, as for any new request. Then use it:
next/font on web, the font
files bundled on Compose and SwiftUI), with font-display: swap on web, behind a fallback
stack matched to it in width, x-height and figure style, named in MASTER.md
("Fira Sans", "Segoe UI", Roboto, system-ui, sans-serif), with size-adjust and the metric
overrides when the fallback shifts the layout.Fira Sans: approved, loaded with a matched fallback, not rendered in this session. The audit hands the check over: the page online, the face loaded, the features
the thesis relies on in place.If python3 is unavailable or the script fails, say so in one line and derive the system from
the thesis by hand. The pipeline does not stop for this.
The MASTER.md design system file is canonical, but the generated code files match the detected stack:
cubic-bezier(...) easings, rem spacing. Output paired with tailwind.config.js extension or :root { --token: ... } CSS.Theme.kt, Color.kt, Type.kt, Shapes.kt, Motion.kt referenced from MASTER.md. Color tokens in Color(0xFF...), typography in TextStyle, shapes in RoundedCornerShape, motion in MotionScheme (M3 Expressive when scope is hero / impactful). Spacing in dp.Color+App.swift, Font+App.swift, Animation+App.swift, Shape+App.swift. Color tokens via Color("AssetName") referencing the asset catalog (or Color(red:green:blue:) if no catalog), typography via Font.system(...) or .custom(...), animations via .spring(...) / .snappy / .bouncy named presets. Spacing in CGFloat constants.commonMain with expect/actual for fonts and platform-specific colors. Same structure as Android Compose, plus a section in MASTER.md describing per-platform deviations.Generate the complete design system based on both theses:
Create a MASTER.md at project root with the full design system. This file is the single source of truth. Every implementation decision references it.
MCP tools, when connected: Stitch (mockups), Nano Banana (visual assets), 21st.dev Magic (components). Without them, carry on: the design system and the code are the core path.
Present the design system in the session's preview mode - announce the mode in one line, don't reopen the menu - and get validation before implementing anything. If Phase 3 chose the display face, show its one-line reason with it: validating the design system validates that amendment to the visual thesis. A palette and a type scale listed as hex codes and pixel values in a transcript are precise and completely unreviewable; every token in MASTER.md is about to be applied everywhere, so this is the cheapest place to catch a wrong one.
Show the dials beside it: each value sent to search.py with the thesis clause it came from,
and each dial left out with the reason.
Load sub-skills based on tech stack and interaction thesis.
Always load (via load_skill <name>, defined above: a missing module prints NOT LOADED and the pipeline carries on, so keep the list for the final report):
load_skill motion-principles - the foundationOne module per shell call. The output of a shell call over about 30,000 characters does not
arrive inline: you get a 2,000-character preview and a file path, and a module you could not
read is a module you did not load. So each loading call re-emits the skill-base block and loads
one module, which keeps its output under 25,000 characters: no module entry file is over that
cap, and validate-skills.py keeps it that way. Keep the tally as you go: load_skill prints
NOT LOADED for a missing module and carries on, and the final report lists both. Load each
module whole: never pipe load_skill or load_ref into head, tail, sed, or grep, since
the batching rule above already keeps each call under the limit, and a truncated module is a
module not loaded.
Context layers (load when applicable):
tells gets a shell call of its own. With its web reference it weighs close to 25,000
characters, too close to the point where a shell call's output stops arriving inline to share a
call with any other module. Re-emit the skill-base block in that call, then:
# (skill-base block re-emitted above this line)
load_skill tells
load_ref tells references/web.md
It loads once the thesis is written and, when someone is there to answer, validated; never
before that, and before any visual choice is frozen: the thesis says which patterns are wanted,
and tells names the reflexes nobody asked for. Never on a Compose or SwiftUI stack, never on
light scope. In paint this row is already met: Phase 3 loads tells before the design-system
query.
Stack-specific (load by SCAN):
"Advanced thesis" trigger for compose-graphics / swiftui-graphics:
The thesis is "advanced" (and triggers loading the graphics sub-skill) if it contains any of these terms:
shader, Metal, AGSL, RuntimeShader, MSLliquid-glass, glassEffect, morphing transitionM3 Expressive, MotionScheme, expressive motioncolorEffect, distortionEffect, layerEffectCanvas (with generative / particle / flow field context)holographic, CRT, displacement, rippleOtherwise stick to the base motion sub-skill.
Read the thesis against the modules, before any code. The thesis (in paint, both) was
written before these modules loaded, so nothing they teach has borne on it yet. After the last
load of this step, and before the variants or the first line of code, walk the rules of each
loaded module that bear on what the thesis promises. A gap is one of two things: the thesis
builds by hand what a module says the platform or the detected library provides (an overlay
drawn by hand where the module starts from the system transition), or the thesis gives up
something a module rule protects (a system gesture, reduced motion, a frame budget). A value or
a device the thesis chose where a module gives a range or a starting point is a decision, not a
gap: the audit check reads it later, on the code. A departure the approved thesis already
names, with what it gives up, is not a gap: the person accepted it. Print one line per gap,
<module>, <rule> - <what the thesis does instead>, or Thesis against modules: no gap, and
go on.
This step runs once. When the revised thesis calls for a module that is not loaded, load it and read the revised thesis against that module alone; a thesis revised here is not revised a second time, and what is left goes to the audit check.
Revised after LOAD: <what changed> - <module, rule>, and build that one. The final report
prints the revised thesis and that line. Keep a departure only when the request asks, in its
own words, for a result the module's way cannot give: quote those words in the thesis, with
what is given up and what replaces it. A request asks for a result: what the person sees and
can do. Words that rule a technique out ("not a push") are read the same way: they ask for
what is seen, so a module's way that gives that look and keeps what the person can do answers
them, whatever the API underneath is called. The technique the thesis chose is not the
request, and a reason that comes from the thesis alone keeps nothing.Shows the thesis before its revision after LOAD, and the final report says so.The audit check "Loaded modules against code" still runs afterwards, on the code.
In paint, a revision made here that changes a value MASTER.md or a token file holds is written into them in the same step, before any page. With someone there, the yes that accepts the revision covers that change.
Implementation rules:
Load the design-audit sub-skill. Phase 4 ended in a user gate, so $SKILL_BASE is gone
again. Re-emit the resolution block in this same call, then:
load_skill design-audit
Run the full audit checklist matching the detected stack. design-audit supplies the greps; the split below decides what you may claim from them.
Before the checklist, outside light scope, read what this run composed from the code, in render order (the markup, the root composable, the root view), and write six labelled entries, each with its file:line:
First screen: what it shows that the visual thesis names there, in the product's own values when it works on values, and treated as the page says the product treats it; that it shows it at rest, from the values the code starts with, before any input, scroll or script; and, from the sizes the code sets, what of it stands whole before any scroll on a short wide frame.Sections: on a wide frame, for the first screen and then each section in order, one line per section: what it adds that no earlier one showed, which starting edge it aligns to, and for each block that is not running text the width it takes and what sets it; two sections that differ carry the thesis's reason, and one that only restates is merged into the one it repeats.Narrowest frame: what the first screen's one thing and every block wider than a line of text become on the narrowest frame the project meets (on native, the compact width at the largest text size), with the rule that does it; a part that carries the idea and leaves the frame, is cut there or loses what names its values is a problem.Empty stretches: every stretch that holds nothing at rest (content that only scroll, script or input brings counts as nothing) and every area left unused beside a block, each with the thesis's reason, or none.Colour roles: each accent and status colour with every job it does; one mark serving both the action and the signal the product exists to show carries the thesis's reason, or the signal gets its own.The one motion: where it lives, or none with the thesis's reason.A promise that comes back not found is built from the thesis; a missing rule is written; an edge, a width, an empty area or a colour job nobody decided gets its decision, stated in that line. One round, on the lines this run wrote, before the checks below. With someone answering, merging a section they validated, or taking a decision the validated thesis does not hold, is proposed in one line before it is applied; only a headless run decides alone. On native, the wide frame is the expanded size class or landscape when the app supports them, otherwise write compact only; a section is a screen of the flow or a block of the root screen; a script is an effect or an async task. In preserve or partial mode, the entries describe what stands and what this run added to it. The six entries open the checked group of the final report and are not counted in its checks; a report asked short keeps them, one line each.
The closing step of this pipeline used to be twenty-five checkboxes, several of which name a tool the agent cannot run. Ticking "60fps verified via Chrome DevTools" without opening Chrome turns "I did not look" into "I looked and it is fine", which is worse than saying nothing.
So the checks are split. Report the first group with the evidence you used. Never tick the second group at all - hand it over.
Each line is reported as check - verdict - the evidence. The evidence is the grep you ran, the
value you computed, or the file:line you read. A verdict with no evidence beside it is not a
finding, and an item you could not check is reported as not checked rather than passed.
file:line that holds it, or
"not found". A thesis that makes no palette or type promise, an interaction thesis alone,
gets no palette or type line. A promise with no file:line is a problem found, not a
pass. Honest downgrade: when promised motion cannot be shipped working, ship the
static version, say so here, and name the promise that was dropped; it counts as a
problem found. Never ship motion that is half broken to keep a promise on paper.
Evidence: one line per promise, e.g. hover 180ms ease-out - src/Card.tsx:42.prefers-reduced-motion block that actually degrades
the animation, not an empty one. SwiftUI: accessibilityReduceMotion. Compose: a helper
on ValueAnimator.areAnimatorsEnabled() / Settings.Global.ANIMATOR_DURATION_SCALE.
Evidence: the file and line of the guard, and what it degrades to.width, height, top, left,
margin or padding; use transform, opacity or graphicsLayer. Evidence: the grep and
its hits, or that it returned nothing.outline: none without a
replacement. Evidence: the grep.file:line.The agent cannot open a profiler, attach to a device, or move a pointer. These are reported as a handoff block with the exact invocation, and marked UNVERIFIED. Do not tick them, do not soften them, and do not omit the section because the rest looked clean.
If a preview or a dev server is already running and the user agrees, driving the browser to collect the web rows is better than handing them over. Never start one just for the audit, and never install anything for it.
Report the two groups separately, with the counts. "11 checked, 2 problems found, 8 handed over" is an honest audit. A single list of ticks is not.
Close the report with the modules, always, as two lines of their own:
Modules loaded: motion-principles, framer-motion, design-audit
Modules not loaded: none
Modules not loaded names every module a load_skill call reported as NOT LOADED, and every
module the load tables called for that was never requested. Write none only when both are
empty. Shell state does not survive between calls, so this list is yours to keep from the first
load to the last: nothing in the shell remembers it for you.
Within the checked group, order the findings by severity: Critical > Important > Nice-to-have. The handed-over group is not ordered and not filtered - it goes over whole, because the user is the one who has to run it.
The final report ends with the feedback door. Once per conversation, on the first final report of a run that wrote code into the project, and never on a run that did not (stopped at a gate, stepped down, handed over to bunshin, modules not found). A headless run prints it too. After the two module lines, print these two lines and nothing more, the link with its values filled in:
How did it come out? One minute, and plain or wrong counts as much as good:
https://github.com/AThevon/genjutsu/issues/new?template=1-output.yml&title=%5Boutput%5D+<skill>&version=<version>&host=<host>&model=<model>
<skill> is cast, paint or bunshin. <version> is the number in the modules from path when it
holds one (genjutsu/<version>/skills/_jutsu), <host> where this ran (Claude+Code+plugin,
npx+skills, claude.ai, Cowork, or the agent's name), <model> your model id. In a value,
write a space as + and percent-encode every other character that is not a letter, a digit,
., - or _. Drop a parameter you do not know, never guess it. Nothing else goes in the URL:
no request, project name, path or content.
Print the link and stop there: never open it, never file the issue and never fill the form yourself, in a headless run either. Nothing is sent unless the user opens the link and submits the form.
When invoked on a project that already has design/styling:
In all three modes, none of these change without the user's explicit agreement, asked item by item: public token names (code outside this project may import them), URLs and routes, navigation labels, form field names, the logo, legal mentions.
Before writing any token file, show the diff between the existing tokens and the new ones
(renamed, changed, added, removed) and wait for approval. After writing, build with the stack's
own command: tsc --noEmit or the project's build script (npm run build or its pnpm / yarn
equivalent) on web, ./gradlew assembleDebug on Android, xcodebuild -scheme <scheme> build on
Apple. When the build cannot run here, hand it over marked UNVERIFIED with the exact command,
like the handed-over group of the audit.
Paint a complete visual universe with genjutsu - art direction brainstorm, design system, implementation, audit. Anti-AI-slop design pipeline. Adapts to Web, Android (Compose), Apple (SwiftUI).
---name: paintdescription: Paint a complete visual universe with genjutsu - art direction brainstorm, design system, implementation…---# skills/paint/SKILL.mdSSS
skills/paint/SKILL.md dosyasını yayıncıdan indirin veya yolu kopyalayıp depoyu açın. PlugMCP beceriyi listeler, dosyayı barındırmaz.
Yayıncı bunu şurada tutuyor: https://github.com/AThevon/genjutsu.
Claude, Cursor, Codex ve Grok Build, yayıncının deposundan eklediğiniz bir SKILL.md dosyasını okuyabilir. Grok Bot, becerinin GitHub klasör bağlantısını bir sohbette gönderdiğinizde onu kurar.
#831843 on #FDF2F8 = 9.4:1div without a role, aria-hidden on decorative motion.
Compose: Modifier.semantics on custom interactive components. SwiftUI:
.accessibilityLabel on controls that have no text. Evidence: the grep.AnimatePresence or the framework's
equivalent; will-change used sparingly and removed after the animation. Evidence: the grep.tells was loaded. In one shell call,
re-emit the skill-base block and run
python3 "$SKILL_BASE/design-audit/scripts/audit.py" . --group tells, whether or not
design-audit was loaded. Evidence, for each finding: either the problem you kept, with its
file:line, or the entry of the thesis's Allowed patterns: line that names it. A kept tell
counts among the problems found. An allowed one is listed as "allowed by the thesis", with
that entry quoted, and is not a problem. The manual reads tells lists (fake product in
divs, repeated layout family, floating corner paragraph, copy register, copy held in
JavaScript data) are reported the same way.
Each finding has one of three outcomes, never a fourth: allowed by the thesis, its
Allowed patterns: entry quoted; not this tell, only when the detection contradicts the
entry's own marker (the loop reports a real loading state, the blur is a modal backdrop, the
matched line is not displayed text), with the file:line that proves it and never on a matter
of taste; or a problem. Fix before you report, and only what this run introduced: a problem
on a line this run wrote (read git diff when the project has one) is removed or rewritten,
then this check runs again, two passes at most, and what still fails is reported with its
file:line. A tell that was already in the project, on a line this run did not write, is
never changed here: it is listed for the user. Nothing on the protected list of an existing
project (public token names, URLs, navigation labels, form field names, logo, legal mentions)
is changed by this step.git diff when the project has one) and report every rule the
code departs from, with its file:line, unless a check above already reported it. A module
rule is a default with its reason, so a departure has two outcomes, never a third:
licensed by the thesis, quoting the sentence of the validated thesis that names the
departure and everything it gives up; or a problem. A sentence that exists only in the
code or in the final report licenses nothing, and neither does a thesis marked UNVALIDATED:
what it states is a problem, reported with the sentence quoted so the user can accept it.
Fix a problem that one local edit removes, then run this check again, once; report the
rest. Evidence: one line per departure, e.g.
mobile-principles, swipe-back - Detail.swift:94 - problem, or the modules walked and
"no departure".