AThevon/genjutsu / skills/paint
paint
VerifiedPaint 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).
Paint - The Master Painter
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.
Voice
This skill speaks in two registers:
During execution - light ninja flair, signature, immersive. Short.
- "Brushing the color palette..."
- "Painting the hero with the unalloyed gold."
- "Setting the spacing tokens."
In reports / final summaries / audit results - plain, factual, dev-readable. Drop the flair entirely.
- "Done. Design system generated. Files: MASTER.md, tokens.css, theme.config.ts. 3 pages painted."
- No mystic prose, no metaphors. Just what changed, files touched, next step.
The flair lives at the intro and during work narration. The moment a result lands or a question gets asked, it's gone.
/paint vs /cast
/genjutsu:cast | /genjutsu:paint | |
|---|---|---|
| Philosophy | "Make this thing beautiful/wow" | "Build a visual universe from scratch" |
| Entry point | Adapts to existing code | Mandatory brainstorm; on an existing design, asks preserve, partial or redesign |
| Discovery | Lightweight, only when vague | Full brainstorm, never skipped |
| Design system | Optional, implicit | Required, generates MASTER.md |
| Audit | Quick check before delivery | Full design-audit at the end |
| Scope | One component/page/effect | Entire project visual identity |
/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.
Iron Rules
- Never skip the brainstorm. Not even if the user says "just make it look good." Especially then. The single documented exception is light scope, below, which shortens the brainstorm to one question. It never removes it. With nobody answering, see "When nobody is answering": the questions are answered from the brief, as assumptions.
- One question at a time during brainstorm. Never bundle. The second question depends on the first answer.
- Never proceed without the theses validated. Visual + interaction, both explicitly approved. The one exception is light scope, below: no visual identity is at stake there, so the interaction thesis alone is required - and it is still validated explicitly, never assumed. With nobody answering, see "When nobody is answering".
- Every design token comes from MASTER.md. No magic numbers, no rogue hex values. On light scope, where no MASTER.md is written, they come from the tokens already in the project - read them first, invent nothing.
- Every animation respects the interaction thesis. Timing, easing, forbidden patterns - no exceptions.
- Never install a dependency without asking. With nobody answering, never install one (see "When nobody is answering").
- Work page by page, validate page by page. Never try to do everything at once.
- The audit is not optional. Phase 5 always runs, even if the user seems happy. On light scope it shortens to the quick check - thesis against code, reduced-motion, exit animation, 60fps - but it never disappears.
- Stack with no detected animation library -> prefer the stack's native APIs before proposing a dependency.
- Animation library detected (GSAP, Motion / Framer Motion, Lottie, Rive, etc.) -> respect the dev's choice. Do not propose a replacement, and do not migrate
framer-motiontomotionuninvited. - Show, don't just describe. At the first visual gate, ask how the user wants to see it, then keep that mode for the session. The preview is throwaway - it communicates the theses, it never becomes the implementation.
Light scope - the one shortened path
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:
- the target is one component, one effect, or one isolated element
- no visual identity is being established: the project already has colors and type, or there is no project yet, only a sketch
- nothing downstream depends on the result being systematised
If two or more fail, it is not light scope. Run the full pipeline and say in one line why.
What changes:
| Phase | Full | Light |
|---|---|---|
| 1 BRAINSTORM | five domains, one question at a time | one question, the least obvious one, then stop |
| 2 THESIS | visual + interaction, both validated | interaction thesis only, still validated |
| 3 DESIGN SYSTEM | generate MASTER.md and the stack token files | skipped. Read the tokens already in the project and use them. Write no MASTER.md. |
| 4 IMPLEMENT | page by page, validate page by page | the one component |
| 5 AUDIT | full design-audit sub-skill | the quick check: thesis against code, reduced-motion, exit animation, 60fps |
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.
Showing Your Work - The Preview Gate
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:
| Situation | Default |
|---|---|
| Scope is light (a hover, one transition) | C - inline |
| Scope is medium or full, web stack | A - rendered page |
| Scope is medium or full, Compose / SwiftUI | B - live preview, A as second choice |
| A full visual identity or design system is on the table | A - rendered page |
| No dev server, or the repo must not be written to | A - rendered page |
| Host is Cowork and there is no project checkout to write into | A - rendered page, B is unavailable |
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:
| Host | A - rendered page | C - inline |
|---|---|---|
| claude.ai | Rendered natively as an artifact. Just produce one. | Written out in the conversation. |
| Cowork | The host's persistent artifact. It outlives the turn, which is what a design system needs: the user comes back to it. | The host's inline widget, rendered in place. Right default for a short task. |
| Claude Code | The Artifact tool when the session has it, else the capability rule below. | Written out in the conversation. |
| unknown (any other host) | The capability rule below. | Written out in the conversation. |
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:
| Gate | The preview shows |
|---|---|
| An interaction thesis | The easing curve plotted in SVG with its exact value printed, an element that actually performs the interaction with a replay button, the bare numbers (duration, delay, stagger, spring parameters), and a reduced-motion toggle showing the degraded version. |
| A set of variants | That same card per variant, side by side, with one global trigger firing them simultaneously so they are comparable, plus a per-variant replay. |
| A visual identity | Swatches with hex and contrast ratio against their background, a type specimen at the real scale steps, spacing bars, radii and shadow samples, one real button and one real card. |
| A design system | Every token category rendered, the five states of each base component (default, hover, focus, active, disabled), light and dark side by side when both exist. |
Rules the preview obeys:
- The message that carries it names the thesis in plain text. Whatever the mode, it opens
with the thesis in one sentence, labelled (
Interaction thesis:orVisual thesis:), then itsAllowed 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. - It is throwaway. It never becomes the implementation. Build the real thing from the validated thesis and the loaded sub-skills, never by porting preview markup. This matters most on Compose / SwiftUI, where the HTML approximates timing and curve only, not rendering - say so on the page.
- Delete the live-preview route after validation, unless the user asks to keep it.
- Never install a dependency to build a preview.
- Never start a dev server without asking.
- Only show values that are in the thesis. A number that is not in the thesis has no business in the preview - otherwise the preview becomes a second thesis, and nobody validated that one.
- A web face the preview cannot load in this session is still shown under its own name, set in its fallback stack and labelled
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.
When nobody is answering
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:
- Discovery and brainstorm questions: do not ask them. Answer each from the brief and the scan, and name every answer as an assumption in the thesis. Read a prohibition in the brief at the scope it states, never wider, and when an assumption takes something out (a picture, a motion, a chosen typeface), say in the thesis what stands in its place.
- Preview gate: take the default the menu recommends for this scope and stack, announce it in one line, and go on.
- Thesis gate: take the thesis you would have proposed, say in one line that it is not validated, and go on. The final report prints it, marked UNVALIDATED.
- Dependencies: never install one. Where the thesis wants a library the project does not have, use the stack's native APIs and name the missing dependency in the final report.
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.
Sub-skills Path Detection
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.
Pipeline
Phase 1 - BRAINSTORM (mandatory, never skip)
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.
Stack scan (run before brainstorm)
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:
- Animation lib: gsap,
motion,framer-motion, three/@react-three, anime.js, or none.motionandframer-motionare the same library at two names. Framer Motion was renamed to Motion;motionis the current package andframer-motionis the legacy one, still widely installed and still published. Note which of the two is inpackage.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. - Framework: React, Vue, Svelte, Next.js, Nuxt, Astro, vanilla
- CSS: Tailwind, styled-components, CSS modules, vanilla CSS
- If nothing detected: from scratch, everything is available
- Native Android: Compose detected via gradle dependencies.
- Native Apple: SwiftUI detected via Package.swift / xcodeproj + swift files. Distinguish iOS vs macOS via Package.swift platforms or pbxproj SDKROOT.
- Compose Multiplatform: kotlin-multiplatform plugin + jetbrains.compose plugin.
- Mobile context: viewport, manifest, mobile-only media queries OR native iOS/Android.
- Desktop context: macOS target OR no mobile indicators on web.
- Legacy mixed: presence of
.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:
- the target is a whole site or web app: several pages or screens, from scratch or as a full redesign;
- the stack is web, or there is no project yet;
- this session can spawn subagents: a subagent tool or a multi-agent workflow tool is among the
tools it exposes (in Claude Code,
AgentorWorkflow);
and at least one of these:
- there is real material to harvest: a social profile, a current site, brand assets, documents;
- the site serves two audiences, or two languages;
- it is a first version meant for a client or a stakeholder;
- the request asks for the best possible, the full treatment, all out.
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: $d"; break; fi
done
The proposal is one message, then wait:
"This is a whole site<, from scratch | , a full redesign>. genjutsu has a pipeline for exactly that:
bunshin. It harvests your real material, researches with parallel agents, lets Impeccable draw the direction (<installed here | not installed: bunshin runs without it, better with it>), builds the pages with parallel agents, and reviews and refines them until only minor issues remain, the round cap of the tier you pick is reached, or a round fixes nothing. It asks you twice: once to start, with a budget, and once to choose the direction. One run has been measured, a seven-page site: about 10.5 million subagent tokens over six to eight hours. The tiers are estimated from it, from about 6 to about 13 million, and scale with the pages. Saybunshinto switch, or I carry on with this pipeline."
- A yes: stop this pipeline and load bunshin. Invoke it as a skill when the host lists
genjutsu:bunshin; otherwise read its entry file next to this one, in$GENJUTSU_SKILL_DIR/../bunshin/(the one Markdown file at the top of that folder), and follow it from its first step. - Anything else: carry on here, and do not propose it again in this session.
- Never when one of the first three conditions fails: a host that cannot spawn subagents cannot run bunshin, and a native stack is not covered by it yet.
- With nobody answering: do not propose and wait. Say in one line that bunshin would fit this brief, carry on with this pipeline, and name bunshin in the final report.
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?"
- Preserve - the theses describe the brand as it is and add only what it lacks (motion, states, missing steps in a scale). Existing tokens stay.
- Partial - the user names what may change ("type and motion, not the colors"). The theses change the values in those areas, under the existing token names, and nothing else.
- Redesign - the theses replace the whole visual layer, existing tokens included.
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:
- Product - What is it (app, landing page, portfolio, SaaS, e-commerce, blog, dashboard...), and what is true of it that its closest neighbour, or a template, could not claim?
- Audience - Who uses it? (devs, designers, general public, enterprise, kids, luxury...)
- Mood - 3 to 5 adjectives that define the visual feel
- References - Sites, screenshots, mood boards the user brings. Never suggest famous ones yourself: that anchors the thesis on someone else's look.
- Tech stack - What's already in place? Or starting from scratch?
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":
- Validate - "That's a starting point. Let's make it precise."
- Go back to the product - "Clean compared to what? Tell me one thing that is true of this product and that its closest competitor could not say about itself." Answer a mood word with a fact about the product, never with another brand's look: "like Stripe, like Linear, like Apple" anchors the thesis on three famous defaults.
- Reframe - "What would feel wrong? What sites make you cringe? That's just as useful."
- Name the consequence - "This choice drives the entire color palette and typography. Worth spending a minute on."
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.
Phase 2 - THESIS (define direction, get validation)
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>.
Visual Thesis
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:
- Color direction - dark/light, palette family, accent color, and the job of each colour: what marks the action, and, when the product has one, what marks the state or signal it exists to show. One mark for both carries its reason
- Typography spirit - the display face by name and why this one, or
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 one - Spacing philosophy - dense/airy, base unit feel, and the starting edges and columns every section answers to: one set, or which sections leave it and why. Reading measure is for running text: a table, a figure or a group of cards says what sets its width
- Component style - rounded/sharp, bordered/filled, elevated/flat
- First screen - of what this run composes (a page, a screen, a window): the one thing it shows, beyond a title and an action, that only this product could show, when the product works on values, in those values rather than marks standing for them, and treated as the page says the product treats it; what it is made of (the project's own material, or an example labelled as an example and named in
Allowed patterns:; an invented customer, metric or quote is a tell, kept only whenAllowed patterns:names it); on a wide frame (on native, the expanded size class or landscape when the app supports them, otherwisecompact 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 aloneis 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 it
Self-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.
Interaction Thesis
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):
- Timing range - fast (100-200ms), medium (200-400ms), or slow (400ms+)
- Hover behavior - what happens on hover
- Scroll behavior - reveals, parallax, or nothing
- Forbidden patterns - what this project will NOT do
- The one motion - the moment where the interface does what the product does, with its trigger, what is on screen before it runs and if it never runs, and its reduced-motion form; or
none, with what states the idea in its place. Not asked on light scope
Cross-platform thesis examples:
- "This Compose hero will use a SharedTransitionLayout with a spring(stiffness=Spring.StiffnessMedium, dampingRatio=0.85) for a fluid card-to-detail transition."
- "This SwiftUI tab transition will use matchedGeometryEffect with a .smooth spring (response: 0.5, dampingFraction: 0.85) for a tactile, spatial feel."
- "This macOS dashboard will use 100ms opacity hover states (no scale on hover, desktop subtlety) and a Cmd+1-9 keyboard shortcut to navigate panels."
- "This Android header will use an AGSL shader bound to scrollOffset for a dynamic liquid-glass effect (Android 13+, with a static fallback below)."
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.
The three dials (after validation)
Once both theses are validated, read three values off them. Phase 3 passes them to search.py
as 1-10 dials:
| Dial | Read it from | 1 | 10 |
|---|---|---|---|
--variance | the visual thesis: layout and component style | centered, symmetric, minimal | bold, asymmetric, experimental |
--motion | the interaction thesis: timing range, scroll behavior, forbidden patterns | subtle | complex, choreographed |
--density | the visual thesis: spacing philosophy | spacious | dense, dashboard |
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.
Phase 3 - DESIGN SYSTEM
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:
- The display face, serif or sans, stays only when the visual thesis names it and says why that face: what it carries for this product that another would not. "Creative", "premium" and "editorial" summon an editorial serif; "clean", "modern" and "SaaS" summon a heavy grotesque. Neither set of words is a reason, and trading one reflex for the other is not a decision. When the thesis left the face undecided, choose it here: name it and say why in one line. That line amends the visual thesis when the design system is validated, and Phase 4 and the audit read the amended sentence, never the original gap.
- A warm paper ground (a faintly warm off-white behind near-black ink and one red accent) is the palette the model reaches for on almost any brief. It stays only when the thesis says why this product belongs on paper.
- Glassmorphism, which the dataset attaches to "SaaS", does not pass unless the thesis asks for it by name.
- Anything else in the output that
tellslists 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:
- Load it the way the stack loads fonts (the Google Fonts URL or
next/fonton web, the font files bundled on Compose and SwiftUI), withfont-display: swapon 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), withsize-adjustand the metric overrides when the fallback shifts the layout. - Check what the thesis needs from it (tabular figures, a weight, a glyph such as the degree sign) in its published data when this session can read it: its specimen page, its OpenType feature list. When it cannot, say which feature is unchecked. An unchecked feature is a reason to tell the user, never a reason to drop the face.
- Say it where the user reads it: in the design-system message, the preview, and the final
report, as
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.
Stack-aware token generation
The MASTER.md design system file is canonical, but the generated code files match the detected stack:
- Web stack detected: generate Tailwind config / CSS variables (existing format). Tokens in CSS hex,
cubic-bezier(...)easings,remspacing. Output paired withtailwind.config.jsextension or:root { --token: ... }CSS. - Android Compose stack detected: generate Kotlin design tokens. Output
Theme.kt,Color.kt,Type.kt,Shapes.kt,Motion.ktreferenced from MASTER.md. Color tokens inColor(0xFF...), typography inTextStyle, shapes inRoundedCornerShape, motion inMotionScheme(M3 Expressive when scope is hero / impactful). Spacing indp. - SwiftUI stack detected (iOS / macOS / multi-target): generate Swift extensions. Output
Color+App.swift,Font+App.swift,Animation+App.swift,Shape+App.swift. Color tokens viaColor("AssetName")referencing the asset catalog (orColor(red:green:blue:)if no catalog), typography viaFont.system(...)or.custom(...), animations via.spring(...)/.snappy/.bouncynamed presets. Spacing inCGFloatconstants. - Compose Multiplatform stack detected: generate Kotlin tokens in
commonMainwithexpect/actualfor fonts and platform-specific colors. Same structure as Android Compose, plus a section in MASTER.md describing per-platform deviations. - Multi-stack project (e.g., web admin + native mobile app): generate MASTER.md with clearly delimited sections for each stack, and produce code files for each.
Generate the complete design system based on both theses:
- Color palette - Primary, secondary, accent, neutrals, semantic (success/warning/error/info). Light + dark if needed.
- Typography - Font stack, size scale (fluid or fixed), weight usage, line-height rules.
- Spacing - Base unit, scale (4px, 8px, 12px, 16px, 24px, 32px, 48px, 64px...).
- Radii - Border radius scale (none, sm, md, lg, full).
- Shadows - Elevation levels (0-4), consistent with visual thesis.
- Base components - Button, input, card, badge, link - styled per the theses.
- Motion tokens - Duration scale (fast/normal/slow), easing names, stagger delay.
MASTER.md
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.
Show it before Phase 4
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.
Phase 4 - IMPLEMENT
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 foundation
One 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):
| Detected | Load |
|---|---|
| Web stack and scope is medium or full | load_skill tells (dedicated call, with references/web.md, see below) |
| Mobile context (web mobile OR native iOS / Android) | load_skill mobile-principles |
| Desktop context (macOS OR web desktop with no mobile indicators) | load_skill desktop-principles |
| Audit explicitly requested OR scope=full | load_skill design-audit |
| Advanced UI/UX questions | load_skill ui-ux-pro-max |
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):
| Detected stack | Sub-skill to load |
|---|---|
| gsap | load_skill gsap |
motion or framer-motion (same library, two package names) | load_skill framer-motion |
| Pure CSS / Tailwind / no lib | load_skill css-native |
| three / @react-three | load_skill threejs-r3f |
| Canvas / generative | load_skill canvas-generative |
| Android Compose | load_skill compose-motion (always) + load_skill compose-graphics (if scope=full or thesis is advanced - see below) |
| Compose Multiplatform | load_skill compose-motion + load_skill compose-multiplatform (always); load_skill swiftui-motion if iOS target detected and SwiftUI interop demanded; load_skill compose-graphics if advanced |
| SwiftUI iOS or macOS | load_skill swiftui-motion (always) + load_skill swiftui-graphics (if scope=full or thesis is advanced) |
"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,ripple
Otherwise 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.
- Someone is there. A thesis a person approved is never changed silently. Give each gap in one line, numbered, with what the module's way keeps and what it cannot do, and ask once: revise or keep, per number. A revise changes only what the gap touches. A keep adds one sentence to the thesis: the departure, what it gives up and what replaces each thing given up. Show the thesis once, with every revision and every keep sentence in it, and wait for a yes.
- Nobody is answering. A thesis the request itself gave or approved stands as approved:
build it, and list each gap in the final report. Otherwise nobody accepted the cost,
so the module's way stands: rewrite the thesis on it, changing only what the gap touches,
print the revised thesis whole, marked UNVALIDATED, with one line under it,
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. - The preview made before LOAD shows the earlier thesis, and nothing is built from it. With
someone there, show the revised thesis again in the session's mode when the revision changes
what the preview showed (the movement, the curve, a number), in one line otherwise. With
nobody answering, do not rebuild it: a preview file still in the project gets a first visible
line,
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:
- Work page by page or component by component - never try to do everything at once.
- Every color, font, spacing, shadow, radius MUST come from MASTER.md tokens. No magic numbers. A colour keeps the job the thesis gave it.
- Every animation MUST respect the interaction thesis (timing, easing, forbidden patterns).
- Apply the 5-state rule for interactive elements: default, hover, focus, active, disabled.
- Every block wider than a line of text (a table, a figure, a row of columns, a long unbroken string) gets its narrowest-frame form in the same edit that writes it, on native at the compact width and the largest text size: what it becomes there is a decision, never what is left over, and each value stays beside what names it (its header, its unit).
- A stretch whose content arrives by scroll, script or input holds its resting content without them. A stretch or an area left empty is a decision the thesis states, or it is closed. A section or screen adds something no earlier one showed; one that only restates is merged into the one it repeats. These rules govern what this run writes: a section that was already in the project is never merged or closed by them.
- Ask the user for validation after each major page/section before moving to the next.
Phase 5 - AUDIT (never skip)
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, ornone.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, ornonewith 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.
Checked here, with evidence
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.
- Thesis against code. First, because every check below assumes the code is the thesis
that was validated. For each promise the validated thesis actually makes (durations,
easing, springs, palette, type, layout family) give the
file:linethat 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 nofile:lineis 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. - Reduced motion honoured. Web: a
prefers-reduced-motionblock that actually degrades the animation, not an empty one. SwiftUI:accessibilityReduceMotion. Compose: a helper onValueAnimator.areAnimatorsEnabled()/Settings.Global.ANIMATOR_DURATION_SCALE. Evidence: the file and line of the guard, and what it degrades to. - Exit animations present wherever something unmounts. Evidence: the conditional render and its exit path, or the list of unmounts that have none.
- No layout-property animation. Nothing animating
width,height,top,left,marginorpadding; use transform, opacity orgraphicsLayer. Evidence: the grep and its hits, or that it returned nothing. - Focus visible on every interactive element, and no
outline: nonewithout a replacement. Evidence: the grep. - All five states on interactive elements: default, hover or press, focus, active, disabled. Evidence: the states you found per component, and the ones missing.
- Tokens, not magic numbers. Colours and spacing come from the project's design tokens
(MASTER.md when one exists). Evidence: the rogue values, with
file:line. - Contrast at least 4.5:1 for body text, 3:1 for large text and UI boundaries.
Compute it from the token values you emitted; do not eyeball a swatch. Evidence: the
pair and the computed ratio, e.g.
#831843 on #FDF2F8 = 9.4:1. - Semantics. Web: no clickable
divwithout a role,aria-hiddenon decorative motion. Compose:Modifier.semanticson custom interactive components. SwiftUI:.accessibilityLabelon controls that have no text. Evidence: the grep. - Web only. Conditional renders wrapped in
AnimatePresenceor the framework's equivalent;will-changeused sparingly and removed after the animation. Evidence: the grep. - Tells confronted with the thesis. Only when
tellswas loaded. In one shell call, re-emit the skill-base block and runpython3 "$SKILL_BASE/design-audit/scripts/audit.py" . --group tells, whether or notdesign-auditwas loaded. Evidence, for each finding: either the problem you kept, with itsfile:line, or the entry of the thesis'sAllowed 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 readstellslists (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, itsAllowed 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 thefile:linethat 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 (readgit diffwhen the project has one) is removed or rewritten, then this check runs again, two passes at most, and what still fails is reported with itsfile: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. - Loaded modules against code. For each module loaded, walk its rules that bear on the
lines this run wrote (read
git diffwhen the project has one) and report every rule the code departs from, with itsfile: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".
You must run these - not verified here
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.
| Target | What to run | Pass condition |
|---|---|---|
| Web | Chrome DevTools > Performance, record across the interaction | no frame over 16.7ms |
| Web | The page at 375 / 768 / 1024 / 1440 | no horizontal scroll, no clipped content |
| Web | The page with the OS "reduce motion" setting on | the degraded path actually runs |
| Web | The page online, when a web face was not rendered in this session | the face loads instead of its fallback, with the figures and weights the thesis relies on |
| Compose | Layout Inspector > Component Tree > View Options > Show Recomposition Counts | counts stable while scrolling |
| Compose | androidx.benchmark.macro Macrobenchmark on a mid-range device | frame time under 16.67ms at 60fps, 8.33ms at 120fps |
| SwiftUI | Instruments > Animation Hitches | no hitch during the transition |
| SwiftUI | Reduce Motion on, Dynamic Type at 200% | nothing clipped, nothing that only moves |
| macOS | Pointer over every interactive element; keyboard through the whole view | hover states fire, focus ring visible, shortcuts bound |
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.
Existing Project Protocol
When invoked on a project that already has design/styling:
- Still run the full BRAINSTORM (Phase 1)
- Ask the mode question in Phase 1: preserve the brand, partial, or redesign. The mode decides what the thesis may replace: nothing already defined in preserve mode, only the areas the user named in partial mode, the whole visual layer in redesign mode
- In Phase 4, existing design tokens are replaced only in redesign mode. In preserve mode, MASTER.md takes the existing tokens as they are and adds the new ones beside them. In partial mode, it keeps every existing token name, changes the values only in the areas the user named, and adds new tokens beside them
- Preserve functionality and layout structure - only replace the visual layer
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.
Red Flags - You're About to Violate This Skill
| Thought | Reality |
|---|---|
| "The user already said 'minimal dark' - I have enough for a thesis" | Two words aren't five domains. Keep asking. |
| "I'll ask all five brainstorm questions at once" | One at a time. The answer to 'audience' changes how you ask about 'mood'. |
| "The user seems impatient, let's skip to coding" | Use the pressure protocol. A bad thesis costs days, not minutes. |
| "This animation would be cool even though the thesis says no bounce" | The thesis is law. Change it? Re-validate with the user first. |
| "The audit can wait, the user seems happy" | The audit is not optional. Phase 5 always runs - shortened on light scope, never skipped. |
| "Nobody is answering, so the safest thesis is the one that promises least" | A thesis that shows nothing passes every check and loses the page. Name the first screen and the one motion, or what stands in their place. |
| "The modules say otherwise, but the thesis is written, the audit will list it" | The audit comes after the code. Read the thesis against the modules right after LOAD: ask, or with nobody there revise it, unless the request approved it. |
| "I can't render the approved font here, the system face is safer" | The sandbox is not the site. Load the approved face behind a matched fallback and say it was not rendered here. |
| "This is a whole site, I'll switch to bunshin, it is clearly better" | Propose it once, with its cost, and switch only on a yes. It spends millions of tokens. |
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 468k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 468k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 468k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |