Shape agent capabilities into your workflow.

agent-or-workflow reads a PRD, ticket, or code sample, extracts seven design signals, and passes them to a deterministic verdict engine. It chooses among a single model call, a workflow, a model-assisted workflow, one tool-using agent, or multiple agents, then writes a decision record with oversight, risks, and the cheaper alternative. The model interprets the design; the script owns the mapping.

fable-setup-optimizer reads Claude Code settings, plugins, agents, and external CLI availability, then presents a done/partial/not-done assessment. With approval before each edit, it can route exploration to cheaper models, load plugins per project, reduce context overhead, tier effort, and delegate typing. Changes to model, plugins, or CLAUDE.md take effect in a later session.

content-audit scans user-facing text across sites and apps for AI-like filler, TODOs, internal implementation details, placeholders, and weak UX copy. It reports each finding with the original line, location, and a drop-in revision, then gives separate AI-likeness and release-readiness scores. The skill evaluates wording and visible leaks, not the truth of external facts.

video-reader drives ffmpeg and ffprobe to extract frames, metadata, spectrograms, and waveforms from local media, with optional Whisper transcription and yt-dlp URL downloads. It exposes both standalone Python scripts and a Claude Code skill, while the core features require only ffmpeg. Frames are downscaled and capped by default so visual inspection does not overwhelm the context.

seo-writer asks for the audience and publishing method, researches competitors, real questions, and authoritative data, then pauses for outline approval before drafting a Traditional Chinese article. It adds direct-answer sections, FAQ, structured data, and SEO metadata, followed by a second quality checkpoint. The result can go through WordPress MCP or be saved as Markdown, and the optional humanizer skill changes the copy style.

goal-forge compiles a task into a /goal contract with explicit deliverables and completion conditions, storing raw command output in a hash-chained ledger. In heavy mode, a prosecutor and three method-diverse jurors re-run checks before a unanimous DONE verdict; small tasks may use one tool-equipped auditor instead while keeping the same evidence and ledger rules. The ledger is tamper-evident, not cryptographic authentication.

3-Layer Synthesis Kit preserves your notes and runs three generative passes: free ideation, forced combinations of atomic fragments, and a cross-pass between the two. Surviving ideas are grouped into four tiers by how strongly they connect to the source notes, with a Mermaid diagram in the output. It proposes possibilities rather than organizing notes or deciding which idea is valid.

publish-to-github-skill creates a new GitHub repository, uploads files through the browser, and checks the result through raw GitHub content rather than trusting the upload screen. It also preflights README, license, and ignore-file hygiene and calls out the executable-bit loss that can affect uploaded scripts. A signed-in Chrome session with Claude browser tools is required; no API key or gh CLI is needed.

frontend-dev-skill creates a project-specific design DNA before coding, then checks the work against a catalog of AI-looking patterns and an append-only design log. It can analyze reference sites, transform sourced components, wire full on-page SEO, and run desktop/mobile screenshot critique. Semrush and browser tools are optional, and the skill chooses among plain HTML, Astro, Next.js, or Nuxt based on the project.

github-viral-marketer packages a repository for promotion by auditing its first-screen hook, generating social previews and SEO tags, and drafting Show HN or Reddit narratives. Its scripts can fork, edit, and submit pull requests to Awesome lists, while an outbound loop finds missing files in target repositories, fixes them, and opens backlinking pull requests. It focuses on repository wrapper, Markdown, and distribution material rather than core code.

human-readable-refactor performs focused, behavior-preserving cleanup over a file, directory, or diff. It removes comments that merely narrate code, improves vague private names, simplifies control flow, and deletes dead or duplicated logic while preserving public APIs and observable error behavior. It favors local conventions and narrow verification, and explicitly does not hide AI use, rewrite authorship, or introduce broad architecture changes.

public-client-leak-audit scopes the public/private boundary, sweeps for backend paths, infrastructure details, rate-limit mechanics, secrets, and tests that reveal server behavior, then checks client-side hardening. It covers permissions, message origins, token storage, DOM sinks, origin pinning, build-time configuration, and remote code or config. The rewrite rule keeps observable client contracts while hiding how the private service works.

Humanizer EN reviews English prose for 36 recurring machine-writing patterns, including empty praise, vague attribution, filler transitions, chatbot leftovers, structural noise, and citation artifacts. It keeps the author’s meaning and does not invent sources, fake imperfections, or rewrite text to evade detectors. The skill can run in Claude.ai, Claude Code, or another compatible agent.

HWPX Skill helps an agent draft Korean administrative documents and output them as HWPX files, following the Ministry of the Interior and Safety’s AI-era writing guidelines. It combines Korean public-sector document drafting with file generation in one workflow for Claude Code users.

Clean Markdown Human runs a Markdown document through focused editorial passes that remove chatbot artifacts, placeholders, filler, vague wording, structural noise, and bad hard-wrapping. It preserves the author’s meaning and does not fabricate sources, dates, features, quotes, or personal details. Code blocks, commands, links, citations, and valid data tables are left alone unless clearly broken.

Baseline is a zero-dependency readiness checker for repositories. It produces a scorecard covering build and tests, security and supply chain, reproducibility, operability, change governance, community and documentation, and claims discipline; blockers can fail CI while judgment calls go to a dated sign-off ledger. It can run as a Claude Code skill or directly through its Node CLI.

Adversarial Design Review convenes specialist agents to challenge a difficult design before code is written. They look for concrete failure scenarios in drivers, on-disk or wire formats, concurrency, protocols, and the CI checks meant to prove correctness; you reconcile the findings and apply fixes while they are still cheap. It is aimed at work that is hard to debug or reverse after implementation.

Translate Ru En converts Russian prose, Markdown, fiction, technical specifications, chats, or directory trees into natural English. It preserves headings, code blocks, links, identifiers, tables, numbers, formatting, and the author’s voice, translating meaning rather than word-for-word calques. Source instructions are treated as text, not executed, and code or technical names stay unchanged.