- name
SKILL.md
The skill file stays in the publisher's repository.
Paste the copied command in a terminal. It installs this SKILL.md into that agent.
npx skills add AThevon/genjutsu --skill orchestration -a cursor -yPaste the copied command in a terminal. It installs this SKILL.md into that agent.
npx skills add AThevon/genjutsu --skill orchestration -a claude-code -yPaste the copied command in a terminal. It installs this SKILL.md into that agent.
npx skills add AThevon/genjutsu --skill orchestration -a codex -yPaste the copied command in a terminal. It installs this SKILL.md into that agent.
npx skills add AThevon/genjutsu --skill orchestration -a grok -ySend the copied message in a Grok Bot chat. The Bot fetches this SKILL.md from GitHub and adds it to the skills all your Bots share.
Install this skill: https://github.com/AThevon/genjutsu/tree/HEAD/skills/_jutsu/orchestrationFrom the publisher
Version-sensitive. The workflow tool contract (
scriptPath,args,agent,parallel,phase) and the Impeccable verbs named here were checked on 2026-10-31. What against is in_jutsu/VERSIONS.md, section "Orchestration". If that date is old, re-check before relying on a verb.
Loaded by
/genjutsu:bunshin. One art director, many clones. This module is how the clones are briefed, fenced, checked and stopped, so that ten agents in parallel produce one site and not ten opinions.
A clone is a subagent. It starts from nothing: no conversation, no memory of the thesis, no idea which files its siblings are editing. Everything it knows is in its brief. Everything that goes wrong in a fan-out goes wrong because a brief left something out.
Each phase that fans out has a template under workflows/, next to this file. Each one is a
workflow script: a pure meta literal, then a body that calls agent(), parallel() and
phase(), parameterised entirely by args. scripts/check-workflows.mjs runs every template
against stubs in CI: to the end with every agent answering, and again with every agent
answering null, where a template must end or stop on its own named error (research.js and
review.js do, with nothing to synthesise or plan), never crash. build.js and refine.js
share one isolated-build block, which the check holds identical in both.
The header comment of each template documents its args. Read it before the call; do not guess
a field.
Running one. When the session has a workflow tool (Claude Code: Workflow), pass the
template by path and its arguments as a JSON value, never as a JSON string:
Workflow({ scriptPath: "<SKILL_BASE>/orchestration/workflows/review.js", args: { ... } })
<SKILL_BASE> is the absolute path the skill-base block printed (genjutsu: modules from ...).
The call returns at once; the result arrives when the workflow finishes. Do not poll it.
Without a workflow tool, with a subagent tool (Claude Code: Agent; other hosts name it
differently): the template is the plan. node "$SKILL_BASE/orchestration/scripts/brief.mjs" <template> <args.json> prints every clone it would spawn, with its label, agent type, schema and
exact prompt, so nothing is rebuilt by hand. Spawn them several in one message when the host runs
them concurrently, one by one otherwise, each schema as the "return exactly this JSON" part.
Calls marked afterAnswers carry a placeholder where an earlier answer goes: rebuild those
prompts with the real answers.
Without either, there are no clones. bunshin does not run: say so and step down to paint.
Never play ten clones in one context: that is one long session reviewing its own work, which
is exactly what the independent lenses exist to avoid.
Return small, write big. A template returns what you need to decide the next step and writes the rest to disk: the full reviews go into the plan file, not into your context. Keep it that way when you adapt one.
A brief has six parts, in this order. A clone that is missing one of them improvises it.
Allowed patterns: line, the tells the thesis does not allow, the voice, the truth rule
(nothing invented, unknown facts generic on the page and marked in the content file), mobile
first, reduced motion. Copy them in: a clone cannot read the conversation where they were
agreed.Add a seventh part when it applies: what is already verified, so reviewers do not spend a lens re-reporting a green build.
Write the brief so it survives being read by a model that has never seen the project. If a sentence only makes sense with the conversation behind it, rewrite it.
Parallel clones on one working tree are safe only when no two of them write the same file.
shared, or to you.shared_change_requests entry: the file, the change, why. You apply them, or you give them to
the shared owner in the next round. Two clones editing the global stylesheet at once is how a
fan-out ends in a merge by hand.shared runs first and alone, then the others run with
its notes (refine.js does this). A page clone that codes against a component prop that does
not exist yet breaks its own build..bunshin/owners.json, every owner included. Pass all of
it to review.js and verdict.js, so the plan can assign any file; a file no owner lists goes
to orchestrator, which is you. Pass build.js only the page owners you did not build
yourself, each with its mission: never shared, never the signature surface. Pass
refine.js only the owners that have fixes, or rules of the decisions file that touch their
files.build.js runs, so a clone's page builds and captures from its first run.Clones build in parallel. A build in the shared working tree reads files other clones are halfway through writing, and every clone chases errors that are not its own.
Each clone builds in a copy of its own, outside the project:
D='<scratch>/build-<owner>'; rm -rf "$D"; mkdir -p "$D"
rsync -a --exclude '/node_modules' --exclude '/dist' --exclude '/.git' --exclude '/.bunshin' \
--exclude '/.claude' --exclude '/.agents' --exclude '/.impeccable' \
--exclude '/.astro' --exclude '/.vercel' --exclude '/.next' --exclude '/.svelte-kit' '<root>/' "$D/"
mkdir -p "$D/node_modules" && find '<root>/node_modules' -mindepth 1 -maxdepth 1 \
! -name .astro ! -name .vite ! -name .cache -exec ln -s {} "$D/node_modules/" \;
cd "$D" && ./node_modules/.bin/<framework> build > "$D/.bunshin-build.log" 2>&1; echo "build exit code: $?"; tail -30 "$D/.bunshin-build.log"
build.js and refine.js generate exactly this, with every path quoted.
node_modules, pnpm build decides the install is foreign and reinstalls through the links
into the real one. The binary just builds. The templates refuse a package manager as cmd.node_modules entry by entry, not as one link. Frameworks keep caches inside it
(Astro writes node_modules/.astro by default): one link to the whole folder makes parallel
clones write into the project's own cache at once. Linked entry by entry, each copy gets its
own cache..astro, .vercel, .next,
.svelte-kit, .output, ...), so a stale output never passes for a fresh build, and a clone
never captures yesterday's page.tail exits with tail's code.<scratch> is an absolute path outside the project, resolved before the call: the session's
scratch directory when the host gives one, else the value of
printf '%s\n' "${TMPDIR:-/tmp}/bunshin-<project>". The templates quote it, so an expression
passed as a path would never be expanded, and they refuse one that is not absolute.Reviewers judge what renders. Before any review or verdict, you produce, and the packet names:
references/evidence.md..bunshin/direction.md), the quality floor:
Impeccable's reference/craft-floor.md, or without it the tells module and its web
reference, by absolute path..bunshin/harvest/facts.md, missing.md and sources.json, by absolute
path: the truth lens holds every visible claim against them..bunshin/direction/, named as the QUALITY BAR card; a decision comp, when one
exists, named as a critique reference only.impeccable detect, the tells
audit, overflow at each width, functional tests. Reviewers do not report these again unless
they can show one is wrong.The lean tier merges truth and tech into truth-tech and drops desktop. Each lens gets
the same packet and never sees the others: a lens that has read another lens agrees with it.
Every finding carries a severity (critical breaks conversion, accessibility or credibility; major hurts the perceived quality; minor is finish), a page, a viewport, the evidence, the issue, one precise fix, and the files.
The plan clone is the only one that sees every review. Its job is judgement, not concatenation:
orchestrator: you apply those yourself, before refine.The plan lands on disk (.bunshin/reviews/round-1.json), next to one file per lens that the lens
wrote itself, and refine.js clones read it from there. Never pass fifty fixes through args.
From round 2 on, verdict.js writes the plan: the finish reviewer gives each remaining point an
id, a priority and an owner, and the script builds the file refine reads.
After a refine round, and after new captures:
verdict.js builds the next plan from its remaining
points, with an empty fixes list when nothing is left.You turn cold-eyes notes into decisions, not into tasks. Write .bunshin/decisions/round-N.md
as rules the next round applies everywhere: the one role of each hero image, where the saturated
colour is allowed and where reading happens on a neutral ground, the type roles, the shape of
controls, the shared contracts that change. Then run the next round against that file. A list of
eight tasks fixes eight spots; a rule fixes the pattern that produced them.
This is the one step you never delegate: deciding what the cold-eyes notes mean is art direction.
The loop is refine, capture, verdict. It stops on a rule, never on a feeling. Read the finish reviewer's disposition first:
recapture: the evidence failed, not the build. Recapture and run the verdict (or the
review) again; nothing the invalid round said binds, and the round does not count. rebuild:
rebuild the named regions yourself, then run a full review again, never a verdict pass; the
round counts. fix and ship go to the rules below.Minor points left after the last round go into the final report. They are not a reason for one more round.
Figures from the one run this pipeline was measured on (a seven-page site in two languages), in subagent tokens as the host reported them. The run: one review, three refine rounds (one from the review, two from decisions files), two verdicts with cold eyes; about 10.5M tokens over six to eight hours, with two human touches. The main session's own tokens were not measured. Per unit: research about 0.6M; about 0.4M per page clone (2.3M for six); a five-lens review with its plan about 1.6M; a refine round from a plan about 1.6M with eight owners; a refine round that applies a decisions file about 1.75M; a verdict with cold eyes about 0.35M; the documenter about 0.2M. Two units were not measured and are estimated from these: a three-lens review with its plan, about 1.1M, and a verdict without cold eyes, about half of one with them. The tiers bunshin offers are estimates built from these units.
Announce the estimate before the first clone, from the page count and the tier, and say where the figures come from. After each workflow, note the agent count and, when the host reports it, the tokens. Never pass a tier's round cap without asking.
Running a team of subagents on one design build: clone briefs, disjoint file ownership, isolated builds, the evidence packet, review lenses, the fix plan, verdicts, cold eyes and the stop rule, with t
---name: orchestrationdescription: Running a team of subagents on one design build: clone briefs, disjoint file ownership, isolated builds, the…---# skills/_jutsu/orchestration/SKILL.mdOverview
Running a team of subagents on one design build: clone briefs, disjoint file ownership, isolated builds, the evidence packet, review lenses, the fix plan, verdicts, cold eyes and the stop rule, with the workflow templates bunshin runs.
FAQ
Download skills/_jutsu/orchestration/SKILL.md from the publisher, or copy the path and open the repository. PlugMCP lists the skill and does not host the file.
The publisher keeps it at https://github.com/AThevon/genjutsu.
Claude, Cursor, Codex, and Grok Build can read a SKILL.md you add from the publisher's repository. Grok Bot installs it when you send the skill's GitHub folder link in a chat.