Ar9av/PaperOrchestra / skills/paper-orchestra
paper-orchestra
VerifiedOrchestrate the full PaperOrchestra (Song et al., 2026, arXiv:2604.05018) five-agent pipeline to turn unstructured research materials (idea, experimental log, LaTeX template, conference guidelines, optional figures) into a submission-ready LaTeX manuscript and compiled PDF. TRIGGER when the user asks to "write a paper from my experiments", "turn this idea and these results into a paper", "generate a conference submission", "run paper-orchestra on X", or otherwise wants the end-to-end paper-writing pipeline. Coordinates the outline-agent, plotting-agent, literature-review-agent, section-writing-agent, and content-refinement-agent skills.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 13.8 KB |
| references/anti-leakage-prompt.md | markdown | 2.5 KB |
| references/host-integration.md | markdown | 7.8 KB |
| references/io-contract.md | markdown | 7.1 KB |
| references/outline-reconciliation.md | markdown | 4.7 KB |
| references/paper-summary.md | markdown | 4.1 KB |
| references/pipeline.md | markdown | 7.2 KB |
| scripts/anti_leakage_check.py | script | 3.3 KB |
| scripts/build_pdf.py | script | 28.3 KB |
| scripts/check_idea_density.py | script | 5.6 KB |
| scripts/check_tex_packages.py | script | 5.9 KB |
| scripts/claim_evidence_gate.py | script | 16.1 KB |
| scripts/diff_outlines.py | script | 3.5 KB |
| scripts/init_workspace.py | script | 4.5 KB |
| scripts/test_claim_evidence_gate.py | script | 5.1 KB |
| scripts/validate_consistency.py | script | 9.7 KB |
| scripts/validate_inputs.py | script | 4.8 KB |
references/pipeline.md
View on GitHub# PaperOrchestra Pipeline
Reference for the orchestrator. Source: arXiv:2604.05018, §4 and Fig. 1.
## The 5 steps
```
┌──────────────────────────────────────────────────┐
│ Inputs: I (idea.md), E (experimental_log.md), │
│ T (template.tex), G (guidelines.md), │
│ F (figures/, optional) │
└────────────────────┬─────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────┐
│ Step 1: Outline Agent (1 LLM call) │
│ outline.json = { │
│ plotting_plan, │
│ intro_related_work_plan, │
│ section_plan │
│ } │
└────────────────────┬─────────────────────────────┘
│
┌────────────┴────────────┐
│ PARALLEL │
▼ ▼
┌──────────────────────────┐ ┌──────────────────────────────┐
│ Step 2: Plotting Agent │ │ Step 3: Lit Review Agent │
│ ~20-30 calls │ │ ~20-30 calls │
│ - few-shot retrieval │ │ - parallel candidate │
│ - visual planning │ │ discovery (10 workers) │
│ - render │ │ - sequential S2 verify │
│ - VLM critique loop │ │ (1 QPS, Levenshtein > 70) │
│ - caption │ │ - dedup by S2 paperId │
│ → figures/*.png │ │ - draft Intro + Related Work │
│ → figures/captions.json │ │ (≥90% citation integration)│
│ │ │ → drafts/intro_relwork.tex │
│ │ │ → refs.bib │
└──────────────┬───────────┘ └──────────────┬──────────────┘
└────────────┬─────────────────┘
▼
┌──────────────────────────────────────────────────┐
│ Step 4: Section Writing Agent (1 LLM call) │
│ ONE single multimodal call: │
│ - extracts numeric values from E │
│ - builds booktabs LaTeX tables │
│ - drafts Abstract / Methodology / Experiments / │
│ Conclusion (preserves Intro + Related Work) │
│ - splices figures from F │
│ → drafts/paper.tex │
└────────────────────┬─────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────┐
│ Step 5: Content Refinement Agent (~5-7 calls) │
│ Loop (≤ ~3 iterations): │
│ 1. simulated reviewer scores current draft │
│ 2. apply revision targeting weaknesses │
│ 3. re-score │
│ 4. accept or revert per halt rules │
│ → final/paper.tex + final/paper.pdf │
└──────────────────────────────────────────────────┘
```
## Parallelism rules
- **Steps 2 and 3 are independent** and have no shared state. They MUST run in
parallel when the host supports it. If not, run Step 3 first because its
wall-time floor is set by Semantic Scholar's 1 QPS verification limit.
- Within Step 2: figure rendering jobs are independent and can be parallelized
per `figure_id`. The VLM critique loop within a single figure is sequential
(render → critique → redraw).
- Within Step 3: candidate **discovery** is parallel (10 concurrent web search
workers in the paper, but the host can use whatever its tool supports).
Candidate **verification** via Semantic Scholar must be sequential at ≤1 QPS
to respect the public API rate limit. See `s2-api-cookbook.md`.
## Step 5 halt rules (verbatim from `content-refinement-agent/references/halt-rules.md`)
The refinement loop accepts a revision iff:
```
ACCEPT if overall_new > overall_prev
ACCEPT if overall_new == overall_prev AND net_subaxis_delta >= 0
REVERT otherwise
```
The loop halts when any of the following becomes true:
1. **Iteration cap reached** (default = 3).
2. **Overall score decreased** vs previous iteration → revert to previous snapshot.
3. **Overall score tied** but net sub-axis change is negative → revert, halt.
4. **No new actionable weaknesses** in the simulated reviewer output.
The "best" snapshot at termination is copied to `workspace/final/`.
## Anti-Leakage prompt
Every LLM call that **writes paper content** (Outline, Lit Review draft,
Section Writing, Refinement) MUST be prefixed with the verbatim Anti-Leakage
Prompt at `anti-leakage-prompt.md`. The Plotting Agent and Caption Generation
calls are exempt because they don't generate paper text.
## Cost budget per agent (from App. B)
| Agent | LLM calls |
|---|---|
| Outline Agent | 1 |
| Plotting Agent | ~20–30 |
| Hybrid Literature Agent | ~20–30 |
| Section Writing Agent | 1 |
| Content Refinement Agent | ~5–7 |
| **Total** | **~60–70** |
Mean wall-clock per paper: ~39.6 minutes (paper Table 7) on parallel
infrastructure. Sequential execution will be 2-3x slower.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deployment | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |