creative-agent-skills

Creative Agent Skills

An open-source Claude Code plugin with focused Agent Skills for natural writing and provider-aware prompt engineering across text, image, video, voice, and music systems.

The project turns a creative brief into a usable prompt package: final prompt, variables, platform settings, constraints, assumptions, reference mappings, and a validation checklist. Provider details are isolated in dated references so the shared workflow remains stable as models change.

اقرأ النسخة العربية.

Install in Claude Code

Add the GitHub repository as a marketplace, install the plugin, then reload it:

/plugin marketplace add GenRamzi/creative-agent-skills
/plugin install creative-agent-skills@genramzi-creative-skills
/reload-plugins

Start with the video workflow:

/creative-agent-skills:video-prompt-engineer
Create one continuous 8-second product shot for an unbranded stainless-steel bottle. Separate the creative prompt from platform settings and include a fallback plan.

Claude can also load a skill automatically when the request matches its description.

See complete examples

The examples/ directory contains compact, reviewable packages for:

• natural-writing revision before and after;
• one-shot AI video production;
• a tightly scoped image edit;
• voice design and delivery direction;
• rubric-based prompt evaluation and regression testing.

These examples demonstrate the output contract without treating model names or product controls as permanent facts.

Skill catalog

What a skill returns

Creative prompt skills use a common delivery contract:

1. A concise interpretation of the brief and stated assumptions.
2. The final copy-paste prompt.
3. Separate platform settings that do not belong in the prose prompt.
4. Variables or reference mappings for reuse.
5. Hard constraints and preservation rules.
6. A short validation checklist and, when useful, one controlled variant or fallback.

Design principles

• Quality over detector games. natural-writing improves voice, specificity, rhythm, factual integrity, and revision quality. It never guarantees that writing will pass an AI detector or misrepresents authorship.
• Provider-aware, not provider-locked. A universal workflow handles the creative brief; dated reference files map it to the selected model.
• Prompt versus settings. Aspect ratio, duration, model ID, resolution, and similar controls remain separate when the product exposes them as settings.
• Preserve before changing. Editing workflows explicitly list what may change, what must remain unchanged, and what must not be added.
• Evaluate behavior. Prompts are treated like versioned application code: test cases and acceptance criteria accompany important workflows.
• No invented capabilities. If a model or interface is unclear, the skill labels the uncertainty and asks for the exact product surface instead of fabricating parameters.

Supported surfaces

The current support matrix and verification dates are in docs/SUPPORT_MATRIX.md. Model names and capabilities change frequently; open an issue when an official source has moved or a platform behavior has drifted.

Validate locally

Only Python's standard library is required:

python scripts/validate_skills.py
python -m unittest discover -s tests -v

If Claude Code is installed, also run its native validator:

claude plugin validate .

The repository also audits every published skill with SkillConform in GitHub Actions, providing an independent conformance and deterministic security-review layer.

Contributing

Contributions are welcome: new provider references, reproducible prompt evaluations, translations, documentation corrections, and focused improvements to existing skills. Read CONTRIBUTING.md and GOVERNANCE.md before opening a pull request.

Provider reference updates must cite an official source and include a Last verified date. Do not submit scraped prompt collections, proprietary prompt leaks, detector-bypass claims, or examples that imitate a living artist or impersonate a real person without permission.

Community and launch resources

• Test one skill and share a sanitized result in the early-adopter feedback issue.
• Use the multi-channel launch kit for Reddit, LinkedIn, X/Bluesky, demo videos, and technical articles.
• Submit provider changes through the dedicated provider-drift issue template with an official source and observation date.

Ecosystem, adoption, and partnerships

Creative Agent Skills and SkillConform form a small open-source stack: this repository packages production workflows, while SkillConform checks their structure, references, permissions, and deterministic security signals in CI.

For the problem statement, target users, differentiators, measurable adoption milestones, partnership paths, and due-diligence links, read the project brief.

The highest-value contributions now are reproducible installation reports, with-skill versus baseline evaluations, sanitized user examples, provider corrections backed by first-party sources, and external marketplace listings.

Status

This is an early public release. The instruction layer is usable and structurally validated, every skill includes Agent Skills evaluation cases with observable assertions, and versioned GitHub releases are generated from the plugin manifest. Broader baselines and community-tested provider variants are the next priorities. See ROADMAP.md.

License

Apache License 2.0. See LICENSE. Product and model names belong to their respective owners; this project is independent and is not endorsed by Anthropic or any referenced provider.