Paper Assembly

Orchestrate the entire paper pipeline end-to-end with state management and checkpointing.

Input

$0 — Paper project directory or paper plan

References

Orchestration patterns and state management: ~/.claude/skills/paper-assembly/references/orchestration-patterns.md

Scripts

Check pipeline completeness

python ~/.claude/skills/paper-assembly/scripts/assembly_checker.py --dir paper/ --output checkpoint.json
python ~/.claude/skills/paper-assembly/scripts/assembly_checker.py --dir paper/ --verbose

Scans paper directory, checks 9 pipeline phases, reports missing artifacts, suggests next steps.

Workflow

Step 1: Assess Current State

Scan the paper directory for existing artifacts
Identify which phases are complete vs pending
Build a dependency graph of remaining work

Step 2: Execute Pipeline Phases

Run phases in dependency order:

Phase	Skill	Input	Output
1. Literature	literature-search, literature-review	Topic	Knowledge base, BibTeX
2. Planning	research-planning	Knowledge base	Paper structure, task list
3. Code	experiment-code	Plan	Training/eval pipeline
4. Experiments	experiment-design	Code	Results JSON/CSV
5. Figures	figure-generation	Results	PNG figures
6. Tables	table-generation	Results	LaTeX tables
7. Writing	paper-writing-section	All above	main.tex sections
8. Citations	citation-management	Draft	references.bib
9. Formatting	latex-formatting	Draft	Formatted LaTeX
10. Compilation	paper-compilation	All	PDF
11. Review	self-review	PDF	Review scores

Step 3: State Propagation

After each phase completes:

Save output artifacts to the paper directory
Propagate results to downstream phases
Update the progress checkpoint file

Step 4: Quality Gates

Before proceeding to the next phase:

Verify all required outputs exist
Check for consistency (e.g., all cited keys in .bib)
Validate figures/tables match experimental results

Step 5: Final Assembly

Merge all sections into main.tex
Verify all \includegraphics files exist
Verify all \cite keys exist in .bib
Compile to PDF
Run self-review for quality check

Orchestration Patterns

Sequential Pipeline (AI-Scientist)

generate_ideas → experiments → writeup → review

Multi-Agent State Broadcasting (AgentLaboratory)

# Propagate results to all downstream agents
set_agent_attr("dataset_code", code)
set_agent_attr("results", results_json)

Copilot Mode (AgentLaboratory)

Human can intervene at any phase boundary for review/correction.

Checkpoint Format

{
  "project": "paper-name",
  "phases_completed": ["literature", "planning", "code"],
  "current_phase": "experiments",
  "artifacts": {
    "literature": "knowledge_base.json",
    "plan": "research_plan.json",
    "code": "experiments/",
    "results": null
  },
  "last_updated": "2024-01-15T10:30:00Z"
}

Rules

Never skip phases — each depends on previous outputs
Save checkpoints after every phase completion
Human review is recommended at phase boundaries
All numbers in the paper must trace to actual experiment logs
Re-run downstream phases if upstream changes

Related Skills

Upstream: all other skills (this is the orchestrator)
Downstream: paper-compilation, self-review
See also: research-planning