skills

43 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:

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. 43 skills: 1 base, 22 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)

Orchestrate end-to-end

Manage the collection

What's in the box

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 43 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.