| name | scriptorium-workflow |
| description | Step-by-step execution guide for the scriptorium 5-phase book production pipeline. Use when: running any phase (1 through 5), orchestrating chapter production, managing parallel review steps (R1+R2+R3 simultaneous), handling recovery after interruption, understanding completion markers, debugging a stalled pipeline, or knowing the exact file inputs and outputs for each step. Covers DAG dependency management and the chapter batch parallelization strategy. |
| license | MIT |
Tech-Editorial: 5-Phase Workflow Execution Guide
How to Start
One-sentence startup — say this to your AI tool:
The source code for [project name] is in [directory path].
Please read QUICK_START.md, then ask me any questions.
If you have no questions, start your work.
Resume after interruption:
Please read checkpoint.md and continue where we left off.
The AI reads QUICK_START.md and runs the full pipeline autonomously.
The sections below are the detailed reference for each phase.
Phase 1 — Outline Finalization
Step 1.1 — Architect Analyzes Source Code
Agent: #1 Architect (general-purpose)
File Pointers IN: source code root, templates/source-map.md, templates/outline.md
Output: outline-draft.md, source-map.md
Completion: <!-- ARCHITECT_DRAFT_COMPLETE --> in outline-draft.md
Step 1.2 — Reader Advocate Reviews
Agent: #2 Reader Advocate (general-purpose)
File Pointers IN: outline-draft.md only
Output: outline-reader-feedback.md
Step 1.3 — Content Reviewer Initial Check
Agent: R3 Content Reviewer (general-purpose)
File Pointers IN: outline-draft.md, templates/style-guide.md
Output: outline-r3-review.md
Step 1.4 — Human Decision Gate
Human reviews all three files → produces outline-final.md. This is a hard checkpoint — do NOT proceed to Phase 2 without human sign-off.
Phase 2 — Shared Resources Build
Agent: #0 Orchestrator (direct execution)
Creates the meta/ directory with 5 files:
| File | Purpose | Type |
|---|
meta/glossary.md | Full-book term definitions | Append-only |
meta/style-guide.md | Writing voice, format rules | Static |
meta/metaphor-registry.md | Used analogies/metaphors log | Append-only |
meta/chapter-summaries.md | Per-chapter summaries (grows as chapters complete) | Append-only |
meta/cross-references.md | Inter-chapter reference table | Append-only |
Phase 3 — Per-Chapter Pipeline (Repeat × N chapters)
Step 3.1 — Research
Agent: #3 Researcher (explore — read-only)
File Pointers IN: chapter spec from outline-final.md, source code files for this chapter, meta/glossary.md
Output: research/chXX-research.md
Completion: <!-- RESEARCH_COMPLETE -->
Constraint: Every claim must cite exact source file + line number. No fabrication.
Step 3.2 — Writing
Agent: #4 Writer (general-purpose)
File Pointers IN: research/chXX-research.md, chapter spec from outline-final.md, meta/style-guide.md, meta/chapter-summaries.md, meta/glossary.md, meta/metaphor-registry.md
Output: drafts/chXX-draft.md
Completion: <!-- DRAFT_COMPLETE -->
Word count target: {{每章目标字数}} ± 20%
Step 3.3 — Three-Way Parallel Review
Run R1, R2, R3 simultaneously in separate agent contexts:
| Reviewer | Context Budget | File Pointers IN | Output |
|---|
| R1 Code | ~37K tokens | draft + source files + source-map | reviews/chXX-r1-code.md |
| R2 Consistency | ~30K tokens | draft + glossary + chapter-summaries + metaphor-registry + cross-refs | reviews/chXX-r2-consistency.md |
| R3 Content | ~18K tokens | draft + style-guide only | reviews/chXX-r3-content.md |
After all 3 complete, Orchestrator merges into reviews/chXX-review.md + <!-- REVIEW_COMPLETE -->
Merge priority: R1 ❌ > R3 ❌ > R2 ⚠️ > others
Step 3.4 — Three-Way Parallel Reader Panel
Run RS, RE, RH simultaneously:
| Reader | Persona | File Pointers IN |
|---|
| RS | CS Junior student | draft only |
| RE | 8yr engineer | draft + source-map |
| RH | Tech hobbyist | draft only |
Orchestrator merges → reader-feedback/chXX-panel.md + <!-- READER_PANEL_COMPLETE -->
Step 3.5 — Post-Chapter Long-Memory Update
Orchestrator executes directly (do NOT delegate):
meta/chapter-summaries.md ← append 200-300 word summary
meta/glossary.md ← append new terms from R2 report
meta/metaphor-registry.md ← append new metaphors from R2 report
meta/cross-references.md ← append new inter-chapter references
Phase 4 — Final Audit
Invoke R1 + R2 + R3 in parallel across the full book:
- R1: validates all code snippets are still accurate against current source
- R2: checks global term/metaphor consistency across all chapters
- R3: final readability and sensitivity scan
Output: final-audit-report.md + revised chapters/chXX.md files with <!-- CHAPTER_FINAL -->
Phase 5 — HTML Publishing
Agent: #11 Bookbinder (general-purpose)
File Pointers IN: all chapters/chXX.md, meta/style-guide.md
Output: publish/ directory (index.html + ch*.html + style.css + assets/)
Completion: <!-- HTML_COMPLETE --> in each HTML file
Recovery Protocol
State is always recoverable from files alone:
- Scan for
<!-- *_COMPLETE --> markers: grep -r "_COMPLETE" . --include="*.md"
- Cross-check with SQL todos table if used
- Resume from the last incomplete step — files without completion markers are safe to regenerate
- Partially written files (no marker) = treat as not started
Parallelization Rules
Within a chapter batch: chapters in the same batch can run Steps 3.1→3.4 simultaneously
Between batches: strictly serial — batch N+1 must wait for all long-memory updates from batch N
Recommended batch size: 2-3 chapters (balance parallelism vs. context coherence)
Failure Handling
| Scenario | Action |
|---|
| Review returns ❌ CRITICAL | Re-invoke Writer with specific ❌ items, then re-run only the reviewer who flagged it |
| Word count off by >20% | Re-invoke Writer with explicit section expansion/trimming instructions |
| Agent crashes without completion marker | Re-invoke with same file pointers — safe to overwrite |
| 3+ consecutive failures on same chapter | Mark as blocked in todos, skip chapter, continue; human review after all others complete |
