Install
$ npx skills add chenzhiyong1994/AIGCREADME
# GitHub Repository: chenzhiyong1994/AIGC
**URL:** https://github.com/chenzhiyong1994/AIGC
**Author:** chenzhiyong1994
**Description:** Reusable AIGC creative skills for image, video, feedback, and music / 面向图片、视频、反馈与音乐创作的可复用 AIGC 能力库
**Homepage:**
**Language:** PowerShell
## Stats
- Stars: 4
- Forks: 0
- Open Issues: 0
- Commits: 4
- Created: 2026-08-17T09:12:37Z
- Updated: 2026-09-05T04:56:42Z
- Pushed: 2026-09-05T04:56:38Z
## README
# AIGC Creative Skills
<p align="center">
<strong>English</strong> · <a href="README.zh-CN.md">简体中文</a>
</p>
> Compile ambiguous creative ideas into executable, reviewable, and reusable generation contracts.
[](LICENSE)
[](https://github.com/chenzhiyong1994/AIGC/commits/main)
[](README.md)
[](README.zh-CN.md)
AIGC Creative Skills is a reusable capability library for creative AI agents. It covers image design, video storyboarding and prompting, evidence-based feedback on real outputs, music creation, and a small set of local analysis tools.
This is not a prompt dump, and it does not try to solve every problem by stacking words such as “cinematic” or “premium.” The project focuses on the parts of creative work that are difficult to reuse: assigning clear control responsibilities, connecting character motion with camera logic, preventing reference contamination, turning real failures into scoped rules, and stopping honestly at the design boundary when no generation tool is available.
## Why this project exists
Generation models keep improving, but reliable output often fails outside the model itself: requirements are not translated into visible acceptance criteria, references compete for the same control, actions lack causes or end states, prompts grow without becoming clearer, and lessons from failed attempts disappear inside one conversation.
This repository offers a lightweight, practical alternative:
- **Define an acceptable result first:** start with purpose, subject, space, action, camera, sound, and end state instead of style-word accumulation.
- **Keep the control budget explainable:** every reference image, motion study, and text constraint has one explicit responsibility.
- **Build in a quality loop:** review the draft with evidence, apply the smallest correction, and recheck only the changed area and its dependencies.
- **Make human feedback reusable:** convert feedback into scoped, stateful, evidence-backed rules without turning one preference into a global default.
- **Never fabricate execution:** record generated files only when they actually exist; otherwise deliver an honest design or prompt package.
- **Load context on demand:** keep long platform guides and specialist rules in references, then load only what the current task needs.
## Capabilities
| Entry point | Best suited for | Default deliverable |
| --- | --- | --- |
| [`image-design`](skills/image-design/SKILL.md) | Character sheets, covers, product images, UI, infographics, brand/editorial visuals, storyboard stills, and boundary frames | Visual direction, reference responsibilities, review summary, and one optimized prompt |
| [`aigc-video`](skills/aigc-video/SKILL.md) | Seedance storyboards, text/image-to-video, extensions, edits, motion studies, and complex action | Director contract, minimal reference set, and final video prompt |
| [`aigc-feedback`](skills/aigc-feedback/SKILL.md) | Reviewing images, prompts, or external renders and correcting failures | Failure diagnosis and correction; local learning only when reusable or requested |
| [`music`](skills/music/SKILL.md) | Singles, EPs/albums, lyrics, style direction, Suno/Udio fields, and audio-reference analysis | Only the requested lyrics, style, structure, or platform fields |
| [`netease-music-*`](skills/netease-music-assistant/SKILL.md) | NetEase Cloud Music CLI setup, available search/playback commands, and preference-aware recommendations | Setup guidance or an executable music-assistant workflow |
Image and video work share a common [review and optimization loop](references/quality-guides/%E5%AE%A1%E6%9F%A5%E4%B8%8E%E4%BC%98%E5%8C%96%E9%97%AD%E7%8E%AF.md). Video workflows add detailed constraints for Seedance reference use, action direction, boundary-frame continuity, and information economy. Music workflows keep lyrics, style, arrangement, and platform-specific fields separate so each layer can evolve independently.
When a user asks for examples—or when the task involves template-heavy work such as UI, infographics, typography posters, brand touchpoints, or product breakdowns—image tasks can query a pinned, text-only snapshot of the [`awesome-gpt-image-2` material library](skills/image-design/references/external-style-library.md). The adapter caches templates and case indexes only, never images. Third-party cases are treated as research references, while final prompts still pass through this project's control and quality loop.
## Quick start
```bash
git clone https://github.com/chenzhiyong1994/AIGC.git
cd AIGC
```
### Agents that support project instructions
Open the repository as a workspace. The root [`AGENTS.md`](AGENTS.md) routes each task to the appropriate `SKILL.md`; complex tasks load specialist references only when the selected Skill requires them.
Start with natural language:
```text
Design a 3:4 product poster for an oriental woody fragrance. Write the prompt first; do not generate the image yet.
```
```text
Turn this 15-second two-person duel into an image-to-video plan for Seedance, keeping attack/defense continuity and the action axis stable.
```
```text
Analyze the describable tempo, instrumentation, and production texture of my audio reference, then create an original song direction without copying its melody.
```
### Other agents and workflows
Point the agent to an explicit entry file:
```text
Read skills/aigc-video/SKILL.md completely, load only the references required by this task, and then complete the following request: …
```
The repository contains readable, adaptable workflow specifications and does not depend on a private orchestration service. Real image generation, video submission, and music generation still depend on tools available in your environment. Codex can use its runtime `imagegen`; other agents can connect an equivalent adapter. Without an adapter, the workflow stops at prompts or design deliverables and does not fabricate results.
## How it works
```mermaid
flowchart LR
A["01 Creative goal"] --> B["02 Select a domain Skill"]
B --> C["03 Load only required references"]
C --> D["04 Produce an internal draft"]
D --> E["05 Evidence-based review"]
E --> F["06 Apply the smallest correction"]
F --> G["07 Deliver one final version"]
G --> H["08 Learn from human feedback"]
H -. "Reuse only scoped rules" .-> C
```
The workflow is intentionally compact. A simple task may produce only one prompt. It expands into a material package only when reference dependencies, cross-shot continuity, or platform constraints genuinely require one.
## Repository layout
```text
.
├── AGENTS.md # Project boundaries and task routing
├── skills/ # Directly usable domain capabilities
│ ├── image-design/
│ ├── aigc-video/
│ ├── aigc-feedback/
│ ├── music/
│ └── netease-music-*/
├── references/ # Cross-Skill style and quality references
├── docs/standards/ # Skill structure and task-archive standards
└── tools/audio/ # Objective audio-reference analysis tools
```
The public repository contains reusable capabilities, standards, and compact evals only. `output/` is for local tasks and must stay out of Git and public releases; the repository validator rejects tracked output files, including force-added files. Untracking preserves local files but does not remove artifacts from older commits. See the [output archive standard](docs/standards/output-archive.md) for storage and history boundaries.
Feedback learning has two layers. The repository tracks only anonymized, reviewed `confirmed` baselines. Personal preferences, client feedback, one-off project evidence, and every `candidate` rule live in Git-ignored `.aigc-local/`. This allows local learning without changing defaults for everyone else. See the [privacy and public-release standard](docs/standards/privacy-and-publication.md) for promotion and publication checks.
## Design principles
1. **Facts before imagination:** if the real artifact cannot be inspected, do not claim visual or audio verification.
2. **Minimum sufficient control:** the number of references follows their responsibilities, not the platform limit.
3. **Close the causal chain before styling:** complex action first resolves space, force, contact, reaction, and end state.
4. **Every rule has a scope:** one subjective comment does not automatically become a universal truth.
5. **Separate creation from execution:** design capabilities remain reusable while external generators can be replaced.
6. **Localized documentation, compatible fields:** project instructions and deliverables may be Chinese-first, while platform fields and essential prompts retain broadly compatible formats.
## Local audio analysis (optional)
[`tools/audio`](tools/audio/README.md) can extract duration, BPM candidates, beats, energy, spectral features, and timbre descriptors from local audio. The music Skill then translates those observations into an original creative direction.
```powershell
python -m venv .venv-audio
.venv-audio\Scripts\pip install -r tools\audio\requirements-audio.txt
.venv-audio\Scripts\python tools\audio\analyze_reference.py path\to\reference.mp3 --out analysis.json
```
Audio decoding depends on a locally available FFmpeg installation. Reference audio should be used only to extract describable features—not to copy protected melodies, lyrics, voices, or complete arrangements.
## Security and privacy
- Never commit `.env` files, cookies, tokens, API keys, private keys, login QR codes, client materials, or runtime logs containing personal paths.
- Never commit `.aigc-local/` or `output/`; they may contain personal preferences, client context, material maps, and execution evidence.
- NetEase-related Skills provide configuration placeholders only. Real `appId` and `privateKey` values must remain in local configuration.
- Before contributing learned experience to public rules, remove identities, exact dates, local paths, private asset names, and identifiable story details; retain only non-identifying evidence types.
- If you discover potentially sensitive information, follow [`SECURITY.md`](SECURITY.md) and report it privately instead of pasting credentials into a public issue.
## Contributing
Run `pwsh -NoProfile -File tools/check.ps1` for the offline repository checks. See [repository validation](docs/standards/validation.md) for prerequisites, focused commands, and the distinction between automated tests and manual behavior scenarios.
Contributions are welcome: verifiable failure cases, platform-rule updates, compact evals, and clearer control methods are especially useful. Read [`CONTRIBUTING.md`](CONTRIBUTING.md) before submitting a change.
Instead of another hundred “universal prompts,” we would rather have one rule that explains why something failed, how to fix it minimally, and where the lesson can safely be reused. If the project saves you an iteration, consider bringing the result back to the community.
## License
Licensed under the [Apache License 2.0](LICENSE). Referenced product names and platform documentation remain the property of their respective owners.
Information
Repository
Language
PowerShell
Created
2026/9/5
Updated
2026/9/5