Orchestrate iterative visual style searches with branch prompts, decision graphs, feedback loops, and final direction selection.
日本語の概要は準備中です。原文の説明を表示しています。
Use when publishing, open-sourcing, exporting, sanitizing, or moving code, agent skills, prompts, templates, fixtures, datasets, workshop assets, or other artifacts from a private repository, vendor/runtime environment, or mixed working directory into a public repository or registry. Builds a provenance inventory, license/security review, explicit allowlist, clean staging package, approval dry run, and fresh public clone/install smoke. Triggers on "open source this", "publish these skills", "make this repo public", "export and sanitize", "подготовь публичный релиз", "выложи скиллы", "опенсорсни", "санитизируй и опубликуй". NOT for ordinary upstream bugfix PRs, vulnerability disclosure, or creating a corp-* department.
インストールする前に、エージェントに与えられる指示の中身を確認できます。
Turn a private, vendor-provided, runtime-generated, or mixed artifact into a public package without leaking secrets, private state, or material that cannot be redistributed.
One pipeline:
intent → owner issue tree → inventory → provenance → license → security/privacy → allowlist → clean package → approval → publish → fresh public verification → maintenance
The skill is allowlist-first. Never copy a whole runtime/private directory and hope denylist cleanup finds everything.
Publishing is an outward mutation. Before creating a public repository, changing visibility, pushing a release, or publishing to a registry:
PACKAGE-READY with no unresolved provenance/license/security blockers;A request to "prepare" or "review" a release authorizes read-only analysis and private/internal artifacts, not public publication.
corp-new; it requires a separate approval dry run.Resolve configuration from the user, project instructions, then defaults:
## Safe Public Release Config
- internal_owner_repo: owner/corp-opensource-or-project
- public_github_owner: owner
- staging_root: /tmp/safe-public-release
- allowed_public_licenses: MIT, Apache-2.0, BSD-2-Clause, BSD-3-Clause
- secret_scanners: gitleaks, trufflehog
- approval_mode: explicit-before-publish
Do not block preparation when optional scanners are unavailable. Record the limitation and keep the release blocked until equivalent manual and repository-native checks are completed. Never claim a scan ran when it did not.
Before extraction or packaging, find or create three scopes in the internal owner repo:
Separate unrelated work:
If the user has an issue-management skill such as manager, use it for the issue tree and cross-repo links.
Use exactly one current status:
DISCOVEREDINVENTORYPROVENANCE-CLEAREDLICENSE-CLEAREDSECURITY-CLEAREDPACKAGE-READYAPPROVEDPUBLISHEDVERIFIEDMAINTAINEDBLOCKED-PROVENANCEBLOCKED-LICENSEBLOCKED-SECURITYBLOCKED-OWNER-APPROVALNO-GO-VENDOR-PROTECTEDNO-GO-MIXED-PRIVATE-DATAEvery blocked status needs one concrete unblock_event: source found, written permission, license clarified, secret removed and rotated, scope reduced, or owner approval.
Record:
Do not create the public repository yet.
Create release-manifest.yaml using the bundled template.
For each candidate artifact record:
Visibility inside an application or runtime does not prove ownership or permission to redistribute.
| Class | Default decision |
|---|---|
| Owner-authored source | Candidate after license/security review |
| Third-party redistributable source | Candidate with preserved notices |
| Generated build artifact | Regenerate from allowlisted source |
| Runtime state, cache, queue, session | Exclude |
| Logs, transcripts, histories | Exclude or separate privacy review |
| Credentials, tokens, cookies, keys | Exclude; rotate if exposed |
| Customer, student, employee, user data | Exclude; aggregate only under a separate privacy owner |
| Vendor-protected or unknown source | Block |
| Mixed bundle | Split before review |
| Symlink, submodule, archive, binary | Inspect target/content and license before allowlist |
Every file in the future public tree must be reachable from an explicit allowlist entry.
For each candidate prove:
Unknown provenance → BLOCKED-PROVENANCE.
Do not publish a "cleaned up" artifact whose origin cannot be explained.
Check:
Attribution is not permission. A missing, custom, or unclear license means BLOCKED-LICENSE until the artifact is excluded or permission is obtained.
The skill does not provide legal advice. When license interpretation changes material risk, record the evidence and route the decision to the appropriate owner or counsel.
Exclude at minimum:
.env*, API keys, OAuth state, cookies, credentials;Run checks on the clean staging tree, not only the source directory:
find . -type l -print
find . -type f -size +10M -print
rg -n --hidden -S '(api[_-]?key|secret|token|password|BEGIN [A-Z ]*PRIVATE KEY|Authorization: Bearer)' .
rg -n --hidden -S '(/Users/|/home/|C:\\Users\\|private-|corp-|localhost:[0-9]{2,5})' .
git diff --check
When approved tools are available:
gitleaks detect --no-git --source .
trufflehog filesystem --no-update .
A scanner passing does not replace manual review. A live secret finding stops the release: remove it, rotate it, document internally, rerun all relevant checks.
Never copy the whole source tree first.
allowlist.txt from the reviewed manifest.Example:
rm -rf "$STAGING_ROOT/package"
mkdir -p "$STAGING_ROOT/package"
rsync -a --files-from=allowlist.txt SOURCE_ROOT/ "$STAGING_ROOT/package/"
cd "$STAGING_ROOT/package"
find . -type f -print | sort
Orphaning a SKILL.md from its references/scripts is a failed package, even when the main file is safe.
The package must contain or link to:
README.md;Do not claim compatibility that the public smoke test has not verified.
Use the bundled checklist. Show the user:
Release: <name>
Target: <public owner/repo or registry>
Allowed artifacts: <count and short list>
Blocked/excluded artifacts: <count and reasons>
Source licenses: <summary>
Security/privacy checks actually run: <commands and results>
Fresh-clone verification plan: <command>
Maintenance owner: <owner>
Open risks: <none or list>
Proposed outward action: create repo | change visibility | push | publish package
Stop and request explicit approval. Continue only for the exact target and artifact set approved.
After approval:
refs owner/repo#N;If the public target is different from the approved dry run, stop and re-approve.
Test as an external user from a fresh directory:
workdir=$(mktemp -d)
git clone --depth 1 https://github.com/OWNER/REPO.git "$workdir/repo"
cd "$workdir/repo"
Run the documented public command:
--help or bounded smoke;Repeat the relevant secret/private-path scans on the fresh clone. PUBLISHED becomes VERIFIED only after this passes.
Write a release evidence card in the internal owner issue:
release_status: VERIFIED
public_url: https://github.com/OWNER/REPO
source_version: ...
manifest: release-manifest.yaml
license_basis: ...
security_checks: ...
publication_commit: ...
fresh_clone_command: ...
fresh_clone_result: PASS
maintenance_owner: ...
next_review: YYYY-MM-DD
After every release or blocked decision, add the new failure mode or rule to this skill or its references.
| Failure | Rule |
|---|---|
| Runtime folder copied wholesale | Allowlist-first clean staging |
| Attribution added, license still unknown | Attribution is not permission; block |
| Main skill copied without references/scripts | Release the complete functional bundle |
| Secret scan green, private path leaked | Manual classification plus private-path scan |
| Works in owner checkout, fails after clone | Fresh public-boundary smoke is mandatory |
| Evaluation mixed with redistribution | Separate issues and risk owners |
| Public repo created before inventory | No outward mutation before PACKAGE-READY and approval |
| Vendor asset visible in UI/runtime | Visibility does not imply redistribution rights |
| Release published with no owner | Maintenance owner and next review are required |
Report only verified facts:
Prepared: <artifact/release>
Status: <state>
Public action: published | not published
Target: <URL or planned target>
Allowed: <count>
Blocked/excluded: <count + reasons>
Checks run: <summary>
Fresh public verification: PASS | NOT RUN | FAILED
Owner issue: <reference>
Next gate: <one concrete action>
まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Orchestrate iterative visual style searches with branch prompts, decision graphs, feedback loops, and final direction selection.
日本語の概要は準備中です。原文の説明を表示しています。
Audits the agent rules in the current folder (AGENTS.md, nested AGENTS.md files) and the skill descriptions the agent sees at start, then reports what to cut, move or rewrite and edits only after the user says yes. Use on «аудит правил», «проверь AGENTS.md», «почему агент не слушается правил», «проверь скиллы».
日本語の概要は準備中です。原文の説明を表示しています。
Робот агента Personal Corp. Агент делает себе голову робота из набора деталей, хранит её в штабе в stack/avatars.json и кладёт в штаб страницу avatar/robot.html, где робот стоит за партой своего штаба и вокруг видны отделы, без сервера и интернета. На уроке по ссылке входа с экрана урока робот входит в живой класс и горит маячком, пока агент работает. Триггеры /avatar, /avatar <ссылка входа>, сделай себе аватар, покажи своего робота, приди на урок.
日本語の概要は準備中です。原文の説明を表示しています。
Use when choosing between engines or models for an agent pipeline and the answer must come from measurement on your own data, not from marketing pages: speech recognition for meeting recordings, the model that turns a transcript into notes, the critic model that checks it. Runs every variant in an isolated container, measures time, cost per hour of input (USD and RUB, with price source and date) and quality (WER/CER and course terms for ASR; code checks plus a judge of a different model for LLM steps), writes results.json/csv and one static analytics page. Triggers on "бенчмарк", "сравни движки", "сколько стоит час записи", "какую модель взять для пайплайна", "benchmark the pipeline", "compare ASR engines".
日本語の概要は準備中です。原文の説明を表示しています。
Use when user asks for Claude Code usage stats, weekly analytics, project activity summary, or wants to see what projects were worked on. Triggers on "аналитика", "статистика claude", "cc stats", "weekly report", "что делал"
日本語の概要は準備中です。原文の説明を表示しています。
Use when needing strategic project analysis from multiple independent expert perspectives. Triggers on business decisions, growth strategy, product direction, competitive analysis, or any situation where diverse C-level opinions reduce blind spots
日本語の概要は準備中です。原文の説明を表示しています。