English | 简体中文
Visit the project website → · 中文主页 · Website maintenance
Explore the workflow, project principles, and quick start on the GitHub Pages website. The homepage source lives in site/ and is deployed automatically from main using GitHub Actions.
Turn AI-assisted writing from a one-off generation into a local-first, reviewable, and reusable content production pipeline.
Content OS is an agent-driven workflow for Chinese-language content creators. It connects topic discovery, research, long-form WeChat writing, human calibration, and optional WeChat article illustrations through explicit stages, artifacts, and review gates.
It is not a black box that takes a topic and publishes automatically. Every important stage leaves a local artifact. Research, voice, and external actions are handled separately, and publishing or paid generation always remains an explicit decision.
Many AI writing failures are workflow failures rather than model failures: trend signals get mixed with evidence, drafts overwrite one another, author voice is reduced to a single prompt, human edits never become reusable knowledge, and publishing lacks a clear safety boundary.
Content OS turns those problems into executable constraints:
- Traceable stage artifacts: every step, from
00_选题卡.mdto05_最终稿.mdand WeChat draft preparation, has a stable input and output. - Research stays separate from expression: aggregated summaries are signals, not evidence; important claims return to primary sources.
- Author voice can compound: representative writing, a style profile, and human feedback calibrate the workflow over time.
- Human edits improve voice: readback compares uploaded and edited text for wording, punctuation, and sentence or paragraph rhythm; visual formatting stays outside the style profile.
- External actions have gates: draft uploads, remote updates, and paid image generation require explicit confirmation.
- Sensitive data stays local by default: working files, credentials, brand references, and private writing samples do not enter Git.
flowchart TD
A["Topic discovery — optional"] --> B["Research"]
B --> C["Draft → Revise → Style refinement → Final draft"]
C --> D["WeChat draft box and human review"]
D --> E["Article illustrations — optional"]
D -. "after manual edits" .-> F["Text-only readback against uploaded baseline"]
F --> G["Author style profile"]
Xiaohongshu adaptation and image production are retired; their instructions are retained in docs/archive/ and excluded from active routing. Xiaohongshu remains a topic-discovery source. Readback is a style-feedback branch, independent of publishing and illustration.
You do not need to run every optional stage. Trend collection and in-article illustrations are explicit choices, while the core writing stages advance through Review Gates instead of running unattended from start to finish.
- WeChat and knowledge-content creators who want AI assistance without flattening their voice.
- Small teams that want stable collaboration points across discovery, research, writing, editing, and publishing.
- Codex and coding-agent users who prefer workflows and artifacts to remain visible inside the repository.
- Creators who care about sources, rollback, and accountable human judgment.
git clone https://github.com/chenzhiyong1994/content-os.git
Set-Location content-os
npm ci
npm testNode.js 22.18+ and PowerShell 7+ are required. npm ci installs the locked HTML parser used for text readback; tests use the Node.js built-in test runner. Dry runs and previews are isolated per article.
- Put two to five representative pieces you genuinely like in
docs/references/style-examples/. - Complete
docs/references/style-profile.mdwith rules supported by those samples or by stable feedback. - To keep a consistent cover character, place a reference image at
assets/brand/reference.png. - For a curated source list, copy
source-watchlist.example.mdto the localsource-watchlist.mdand fill it in.
Private samples and brand material are ignored by default or represented by sanitized templates. See the README in each corresponding directory for details.
Open the repository in Codex and ask, for example:
Use the content-workflow to help me create a practical article from topic discovery.
The target audience is everyday office workers.
The Agent reads AGENTS.md, the workflow rules, and the relevant skill before stopping at each required Review Gate. The canonical stage definitions live in docs/operations/workflow-rules.md.
| Capability | When it is needed | Default boundary |
|---|---|---|
| External writing engine (optional) | Delegate heavy writing stages to another CLI | Configure the command and model explicitly; otherwise the current Agent writes |
| Browser access | Validate discussions on signed-in social platforms | Run source health checks first; public search cannot impersonate signed-in scanning |
| ImageGen | Covers and WeChat article illustrations | Show the tool path, visible model information, dimensions, output path, and complete prompt before confirmation |
| WeChat Official Account API | Upload and read back drafts | Supply credentials only through environment variables; dry run is supported |
See .env.example and docs/operations/wechat-publish-setup.md for environment variables. Credentials are never written to logs or the repository.
.
├─ AGENTS.md # Project constitution and task routing
├─ docs/
│ ├─ operations/ # Workflow, feedback, publishing, and version safety
│ ├─ archive/ # Retired instructions, excluded from active routing
│ └─ references/ # Content type, facts, style, and platform standards
├─ skills/ # Executable instructions for each content stage
├─ scripts/ # Local helper scripts
├─ tests/ # Publishing and browser contract tests
├─ site/ # Static project website, deployed to GitHub Pages
├─ assets/brand/ # Your local brand reference material
└─ workspace/ # Temporary files and article outputs, ignored by default
- Evidence before expression: trend signals, research evidence, and author judgment carry different weights.
- Save before review: every stable stage is written to disk before human review.
- Few hard rules, concrete examples: style comes from real writing and feedback, not an ever-growing blacklist.
- Optional still means explicit: optional stages are deliberately chosen, never silently skipped or triggered.
- Local is the source of truth: external platforms distribute and calibrate; they do not replace traceable local source material.
workspace/,.env*, private writing samples, and brand reference images are ignored by default.- WeChat credentials are read only from environment variables such as
WECHAT_MP_APP_IDandWECHAT_MP_APP_SECRET. - Public repository content consists of sanitized templates rather than real credentials, historical articles, author profiles, or brand originals.
- Author names, curated source lists, AI aggregation endpoints, and external writing engines have no maintainer-specific defaults.
- Before publishing your own derivative repository, scan both the current files and Git history. Deleting a file does not remove it from history.
Report security issues through GitHub private vulnerability reporting as described in SECURITY.md. Do not paste credentials or private material into a public Issue.
Contributions that make the workflow more reliable, explainable, or portable are welcome, including:
- New platform adapters and verifiable publishing workflows
- Narrower, more reliable fact checks and regression tests
- Style calibration methods that do not depend on one specific author
- Better Windows, macOS, and Linux compatibility
Read CONTRIBUTING.md before submitting a change. If you are trying the project for the first time, opening an Issue about friction encountered on a real topic is equally valuable.
MIT © Content OS contributors