skills
42 skills for Claude Code that make content — and refuse to let it read like a machine made it.
Prose editing that strips the tells. Prompt engineering for 40+ image, video and
music models, with optional one-command execution against the real APIs. And
orchestrators that carry a job the whole way: research question → carousel →
published post; raw price list → a proposal styled from the client's own website.
Russian-first, English throughout. Plain markdown, MIT, no required dependencies.
What it actually does
The same cold email, before and after cold-email → writer:
Subject: Quick question
Hi Patricia,
I hope this email finds you well! My name is David Kim and I'm the co-founder
and CEO of Nimbus, a B2B SaaS startup that's been disrupting the
developer-tools space for the past 22 months.
I've been a longtime admirer of your work at Magnet Ventures, and I was
incredibly impressed by your recent investment in DevForge. Given the
synergies I see between our work and your portfolio, I wanted to reach out
personally to introduce myself.
We've been growing rapidly — we now have a number of enterprise customers and
significant ARR. […]
Subject: $2.1M ARR in dev tools — fit for Magnet?
Hi Patricia,
Saw the DevForge round — congrats on leading. We're in an adjacent wedge:
pre-merge code-review automation for engineering teams of 20-200. $2.1M ARR
over 22 months, 12 enterprise customers including Stripe, Datadog, and
Cloudflare.
Raising A in Q3. Worth 20 min next week to see if Magnet would be a fit?
Happy to send the deck cold either way.
Best,
David Kim
Co-founder, Nimbus · nimbus.dev
Every cut is attributable to a rule, not to taste — "I hope this email finds you
well" to banned-patterns.md §Ceremony openers, the 24-word bio to
structure.md §Block 1. The full pair with every delta
is in the repo, along with three more.
There is also an offline linter, so the same rules can gate a commit with no API
call and no model in the loop:
$ python3 ~/.claude/skills/writer/scripts/lint.py draft.md
writer-lint: borderline (3 hits)
gate passed: no hard bans.
Hits:
L1:42 [caution] GPT_FILLER "it's important to note"
L2:30 [caution] GPT_FILLER "Let's dive into"
L6:1 [caution] AI_BRIDGE 'Moreover'
Exit codes are hook-shaped, so style-check can run it on staged files and stop
the commit — pre-commit walkthrough.
How it fits together
One skill does the prose work — writer — and eleven others
route through it rather than restating its rules. viral-text knows about hooks
and platforms; it does not know about em-dashes, because writer does. The same
shape holds for media: nine skills build on image-prompt.
That is the whole architecture. 42 skills: 1 base, 21 wrappers, 3 linters,
14 orchestrators, 3 meta.
Install
# Plugin marketplace — inside Claude Code
/plugin marketplace add Mikefluff/skills
/plugin install mikefluff-skills@mikefluff
# Curl — 5 seconds, no deps
curl -fsSL https://raw.githubusercontent.com/Mikefluff/skills/main/install.sh | bash
The plugin route keeps the skills updatable from inside Claude Code
(/plugin marketplace update); the curl route drops them straight into
~/.claude/skills/. Either works — don't run both, or every skill appears twice.
Other methods (npm, Homebrew, Docker, local checkout, pinned version, custom prefix): docs/INSTALL.md.
Skills appear after Claude Code session restart. Discovery is automatic via the name: and description: fields in each skill's frontmatter — no ~/.claude/settings.json edits required.
API keys (for --execute mode — entirely optional): manage via /skills-keys add OPENAI_API_KEY ... inside Claude Code, or copy .env.example to ~/.skills.env. See Managing keys below.
Documentation
The full doc hub: docs/. Most-used pages:
The bundled style libraries used by the orchestrators: carousel · video · music · visual-prompt styles (extensible, 23 entries).
The shared LLM prompt chains behind the visual skills: image chain (carousel / cover / quote / meme / banner / logo — typographic template + style fields + orientation lock) and video chain (reel shots / carousel animation — overlay-heavy i2v discipline, lock_first_last + negative_prompt text-stability). Saved brand profiles for proposal-maker: proposal-maker/brands/.
What you can do
Write & edit prose
Generate AI media (prompt-first, execute optional)
| Scenario | Skill(s) | Walkthrough |
|---|
| Generate an image prompt for 14+ models (Midjourney V8.1, Flux 2, Nano Banana 2, Nano Banana Pro, gpt-image-2, Ideogram 3, Seedream 5.0, …) | image-prompt | image-prompt-cover |
| Generate a video prompt for 20+ models (Veo 3.1 + audio, Sora 2, Kling 3.0, Runway Gen-4 / Aleph V2V, Luma Ray 3, Pika, Hailuo, Hunyuan, Wan 2.2, Seedance, …) | video-prompt | video-prompt-reel |
| Generate a music prompt for 10+ models (Suno v5.5, Udio v4, Lyria 3 Pro, ElevenLabs Music, Stable Audio 2.5, MusicGen, …) | music-prompt | USER-GUIDE |
| Voiceover / narration via TTS (ElevenLabs multilingual + OpenAI gpt-4o-mini-tts) | voiceover-maker | skills/voiceover-maker/examples/before-after.md |
| Burn captions onto an existing video (SRT / VTT / plain text + ffmpeg) | subtitle-burner | skills/subtitle-burner/examples/before-after.md |
| Actually run the prompt through the vendor API and save a real PNG / MP4 / MP3 | any of the prompt skills + --execute | execute-end-to-end |
Orchestrate end-to-end
Manage the collection
What's in the box
| Skill | Layer | Languages | Purpose |
|---|
writer | base | ru/en | Base clean-prose editor — antinyeyroslop (25 catalogued categories + copy-paste artifacts + rhythm metrics), typography, structural synthetics, RU calques. Invoked by all other prose skills. |
viral-text | wrapper | ru/en | Write viral social media content — hooks, numbered points, micro-conclusion with NLP question, CTA. 41 viral content rules + platform adaptation. |
prose-edit | wrapper | ru | Fiction rewrite layer — Pelevin/Manson voice vector, 10-item style drift checklist, no meta-refs / anglicisms in narrator voice, long artistic rewrite (no comma-stitching), ToV pattern, 5-trigger structural-synthesis detector, Postirony depth-pass. |
essay-write | wrapper | ru | Non-fiction layer — long subordinate sentences (Manson style), source-backed claims, philosophy through humor, biography through scenes, plain-Russian for complex content. |
style-check | linter | ru/en | Read-only pre-commit lint that stacks writer + prose-edit + essay-write rules. Routes by configurable path patterns (fiction vs non-fic), BLOCKING/WARNING/INFO severity, exit-code semantics for git hook. |
translation-sync | linter | ru/en/pt-br | Read-only pre-commit parity checker for trilingual book translations (RU↔EN↔PT-BR) — typography per language, terminology canon, anchor-quote drift, names/patronymics/diminutives, cultural realia footnotes, no-smoothing of numbers/brands/dates. BLOCKING/WARNING/INFO severity with exit-code semantics for git hook. |
canon-check |
Skills compose: wrappers call writer internally; linters reference the same rule files but don't mutate; orchestrators chain multiple wrappers + the execute runner. See docs/COMPOSING.md for the full dependency graph and named workflows.
The --execute layer (optional)
image-prompt, video-prompt, music-prompt produce paste-ready prompts by default. Pass --execute and they ALSO call the vendor API and save real PNG / MP4 / MP3 assets. Without API keys they stay prompt-only — no breakage.
carousel-builder and reel-builder are orchestrators built on this layer — they assemble batches of prompts and execute them with a single style anchor + cost confirmation.
# image
~/.claude/skills/image-prompt/scripts/run.py --execute --model gpt-image-2 --prompt "..."
# video (cost confirmation prompts; --yes to skip)
~/.claude/skills/video-prompt/scripts/run.py --execute --model veo-3-1-fast --prompt "..." --duration 8
# music
~/.claude/skills/music-prompt/scripts/run.py --execute --model suno-v5-5 --prompt "..." --lyrics-file ./lyrics.txt
Vendors covered: OpenAI (gpt-image-2, Whisper + GPT-4o transcribe, TTS, Sora 2 until its 2026-09-24 removal), Google (Nano Banana Pro / 2 / 2 Lite, Veo 3.1 + Fast + Lite, Lyria 3 Pro), Black Forest Labs (Flux family + Kontext), Runway (Gen-4, Gen-4.5, Aleph), Kuaishou (Kling 3.0), Ideogram (3 Flash/Turbo/Default/Quality), ElevenLabs (Music + TTS on Eleven v3), Suno (v5.5), fal.ai + Replicate routers for everything else (Seedream 5, Hunyuan, LTX-2, Wan, MusicGen, Stable Audio, many open-source).
Output: assets land at ./generated/<modality>/<timestamp>-<model>.<ext>. If S3_BUCKET is set, they also upload to S3-compatible storage (AWS S3 / DigitalOcean Spaces / Cloudflare R2 / MinIO) and the URL is printed alongside.
Setup is one command:
# Inside Claude Code, after install
/skills-keys add OPENAI_API_KEY sk-proj-...
/skills-keys add GEMINI_API_KEY AIza...
/skills-keys verify # ping each vendor; confirm valid/invalid/unknown
The runner auto-creates ~/.claude/skills/.runners-venv and installs Python deps (requires Python ≥ 3.10). Override with SKILLS_SKIP_VENV=1 if Python is missing.
Full provider matrix + cost preview + per-vendor troubleshooting: skills/image-prompt/references/execute.md, skills/video-prompt/references/execute.md, skills/music-prompt/references/execute.md.
Managing API keys
Keys for the --execute layer live in ~/.skills.env (chmod 600). The runner auto-loads it into os.environ at every CLI startup — explicit shell exports still win.
/skills-keys list # masked overview
/skills-keys add OPENAI_API_KEY # interactive (silent stdin prompt)
/skills-keys update OPENAI_API_KEY sk-proj-new... # rotate
/skills-keys remove OPENAI_API_KEY
/skills-keys enable SUNO_API_ENABLED # gate-flag shortcut
/skills-keys verify # ping 9 providers (OpenAI/Gemini/Anthropic/BFL/Ideogram/Replicate/FAL/Runway/Eleven)
/skills-keys export # eval-ready lines for current shell
Full reference: skills/skills-keys/references/usage.md · examples.
Don't have an API key for one of these vendors? The matching skill stays in prompt-only mode — paste the generated prompt into the vendor's UI manually. No skill requires any key.
Updating
/skills-update # in-Claude check + apply
bash install.sh --check # CLI version diff
bash install.sh --update # re-install from latest release
Optional ambient banner in your shell status line (opt-in): bash scripts/install-hook.sh. Shows · skills v2.3.0→2.3.1 +1 skill when an update exists. Never updates without confirmation.
Common install flags
# Install a subset
curl -fsSL .../install.sh | bash -s -- --skills writer,viral-text
# Custom prefix
curl -fsSL .../install.sh | bash -s -- --prefix /tmp/skills
# Pin a version
curl -fsSL .../install.sh | bash -s -- --version 2.3.0
# Update existing install (re-install + new deps)
bash install.sh --update
# Diff local vs latest
bash install.sh --check
# Uninstall everything
bash install.sh --uninstall
Full installer help: bash install.sh --help · docs/INSTALL.md.
Repo layout
skills/
├── README.md # this file
├── VERSION # semver, single source of truth
├── CHANGELOG.md # Keep-a-Changelog
├── skills.json # machine-readable manifest used by the installer
├── install.sh # pure-bash installer, curl-pipeable
├── .env.example # template for ~/.skills.env (manage via /skills-keys)
├── Makefile # local dev convenience
├── CONTRIBUTING.md # how to add / report / propose a skill
├── CODE_OF_CONDUCT.md # short; scales up if the project does
├── SECURITY.md # how to report a vulnerability
│
├── skills/ # the 42 skills, one folder each
│ └── <skill-name>/
│ ├── SKILL.md # frontmatter + the rules; this IS the skill
│ ├── references/ # deep material loaded on demand
│ ├── examples/ # before/after calibration pairs
│ └── scripts/ # only where a skill ships code (writer's linter)
│
├── common/
│ ├── references/ # shared anti-pattern catalogues (hype words, preambles, …)
│ ├── runners/ # optional Python execute layer (image/video/music)
│ │ ├── providers/ # 14 image + 10 video + 5 music + 3 audio = 32 providers
│ │ ├── cli/ # per-modality CLI entries (image / video / music / carousel / reel / keys)
│ │ ├── styles.py # style library loader
│ │ ├── batch.py # parallel executor for carousel / reel
│ │ ├── ffmpeg.py # concat / audio mix / caption burn-in
│ │ ├── keysfile.py # CRUD over ~/.skills.env
│ │ ├── verify.py # HTTP probes for 9 providers
│ │ └── requirements.txt
│ └── style-library/ # 48 bundled styles (24 carousel + 12 director + 12 music)
│
├── docs/
│ ├── QUICKSTART.md # 5-minute first run
│ ├── USER-GUIDE.md # scenarios with TOC
│ ├── COMPOSING.md # dependency graph + named workflows
│ ├── INSTALL.md # detailed install methods
│ ├── SKILL-INDEX.md # auto-generated, by layer/domain/language
│ ├── FAQ.md # short Q&A
│ ├── TROUBLESHOOTING.md # when things break
│ ├── VERSIONING.md # semver policy + release flow
│ ├── LINTER-COVERAGE.md # auto-generated regex coverage
│ ├── LAUNCH-POST.md # frozen v1.9 launch copy
│ └── walkthroughs/
│ ├── README.md # categorized index
│ └── *.md # 20 detailed flows
│
├── scripts/ # validate + smoke + doc generators
├── hooks/ # ambient update banner (opt-in)
├── tests/ # fixtures + snapshots, runner unit tests, model-in-loop evals
├── bin/ # the `skills` CLI shim (npm package)
└── .github/ # workflows + issue/PR templates
Local development
make help # list all targets
make install # install from this checkout into ~/.claude/skills/
make smoke # 14 gates: validate, linter, snapshots, samples, links,
# unit tests, pricing, markdownlint, coverage, tweets,
# launch copy, imports, code quality, CLI surface
make test-unit # runner unit tests alone
make check-docs # docs-consistency gate
make gen # regenerate every derived file (README table, indexes, coverage)
make new-skill NAME=foo-bar DESC="..."
Releases are manual: make bump-minor → fill the CHANGELOG section → make release (verifies, tags, pushes) → publish the GitHub release. That last step matters — install.sh resolves latest from the newest published release, so a bare tag leaves curl installs on the previous version. See docs/VERSIONING.md.
Contributions welcome — see CONTRIBUTING.md.
License
MIT — see LICENSE.
