writing-workbench/WORKBENCH.md

130 lines
3.8 KiB
Markdown

# WORKBENCH.md
## 1. Workspace Identity
This project is the article production workspace.
It is a stage and process black box for deep writing. It records how an article moves from selected idea to final published text, while keeping the original knowledge archive and CCPE asset factory separate.
## 2. Operating Principle
Preserve process as a first-class artifact.
The user values not only final drafts but also the discussion, review, synthesis, repair, and decision trail that produced them. Do not collapse multi-round writing or review work into a single final file unless explicitly asked.
## 3. Project Lifecycle
Ideas normally begin outside this workspace:
```text
knowledge-vault/diary
knowledge-vault/discussions
knowledge-vault/sources
knowledge-vault/sayings
knowledge-vault/rules
```
A project enters Writing Workbench only after the user decides to write.
Each active article should use a stable project id:
```text
projects/active/YYYY-MM-DD-aNNN-temp-slug/
```
Do not rename the project folder every time the article title changes. Store title changes in `project.md` and final title metadata in `publish/final.md`.
## 4. Canonical Article Project Structure
Use `templates/article-project/PROJECT_STRUCTURE.md` as the expansion guide when creating a full project.
Top-level project files:
- `project.md`: stable project identity, temporary title, final title when known, phase, status
- `source-map.md`: source files from Knowledge Vault and other repositories
- `status.md`: current phase, open decisions, next actions
Main stages:
- `00-creation-plan/`
- `01-positioning/`
- `02-outline/`
- `03-drafting/`
- `04-draft-review/`
- `05-appendix/`
- `06-metadata/`
- `07-tribunal/`
- `08-model-writing/`
- `publish/`
## 5. Session And Thread Policy
Use files as shared state.
Separate Codex sessions or background threads may be used for writers and reviewers, but their long conversation history should not be copied wholesale back into the orchestrator session.
Preferred flow:
```text
orchestrator session
prepares context-pack
dispatches writer or reviewer session
receives session.md and report.md
reads only the necessary report/synthesis
asks the user for decision when needed
```
For each review role, preserve:
- `session.md`: discussion/process record
- `report.md`: final review result
The orchestrator should consume reports and synthesis files, not every full session transcript.
## 6. Context Pack Policy
Do not feed every source file by default.
Prepare task-specific context packs:
- `context-pack-lite.md`: article goal, current outline/draft, core claims
- `context-pack-standard.md`: lite pack plus selected material summaries and key source links
- `context-pack-deep.md`: standard pack plus long source excerpts, historical essays, or Deep Research
Use the smallest pack that lets the role do useful work.
## 7. CCPE Consumption
Writing Workbench consumes CCPE assets from:
```text
C:\Users\wangq\Documents\Codex\ccpe-system
```
Examples:
- main writer agents
- Zhang Liao red-team agent
- Cognitive Imaging
- Giant Cognition
- metadata agent
- context-pack builder skill
- review synthesis skill
- repair request builder skill
- Article Production Orchestrator
- Outline Review Runtime
Do not duplicate the full canonical CCPE asset inside this workspace. Add thin notes in `ccpe-consumption/` that point to the canonical source and explain how this workspace uses it.
## 8. Archive Policy
When an article is completed:
- keep full process records in `projects/completed/`
- archive final essay in `knowledge-vault/sayings`
- archive human model documents in `knowledge-vault/rules`
- promote AI-facing model cards or runtime assets through `ccpe-system` only when mature enough
When a project is abandoned, move it to `projects/abandoned/` and preserve the reason in `status.md`.