Livepaper

Publish research papers as interactive web pages where every number is linked to a replication prompt. Readers click a value, copy the prompt into any AI coding agent, and reproduce the result.

How it works

A livepaper has two source files:

• paper.md — your paper in markdown, with markers on quantitative values
• specs.yaml — replication prompts for each marked value

Markers look like standard markdown links:

Our model achieves [98.8%](#=clf_lr_acc) accuracy on the benchmark.

In the built page, "98.8%" is highlighted. Click it, get a popover with the replication prompt and a copy button. Paste into Claude Code, Cursor, Codex, or any agent.

Values not yet verified use pending markers:

Training took [3 days](#=) on 8 A100s.

Install

npm install -g livepaper-cli

Or use without installing:

npx livepaper-cli <command>

The CLI command is still livepaper.

Workflow

1. Initialize from a LaTeX paper

livepaper init paper/paper.tex

Creates livepaper/ with INSTRUCTION.md — an agent playbook for converting your paper.

2. Let the agent do the work

Open your AI coding agent and say:

Please follow livepaper/INSTRUCTION.md to build my livepaper.

The agent will:

• Convert your LaTeX to livepaper/paper.md with markers on every number
• Explore your repo to find how each result was produced
• Fill in livepaper/specs.yaml with replication prompts
• Ask you when it can't figure something out

3. Build the interactive page

livepaper build

Open livepaper/dist/index.html in your browser.

Commands

livepaper init <tex>

Scaffold livepaper/ and generate INSTRUCTION.md.

livepaper init paper/paper.tex
livepaper init paper/paper.tex --force  # overwrite existing

livepaper build

Build the interactive page from paper.md + specs.yaml.

livepaper build                    # outputs to livepaper/dist/
livepaper build --out ./public     # custom output dir
livepaper build --watch            # rebuild on changes

livepaper check

Validate markers against specs without building.

livepaper check

✓ 25 markers found in livepaper/paper.md
✓ 25 specs found in livepaper/specs.yaml
✓ All markers have matching specs
✓ No orphan specs
✓ All IDs unique
✓ All {var} references resolved
· 7 pending markers (#=)

25/25 ready. 7 pending.

Authoring format

Markers in paper.md

[value](#=id)    verified — has a spec in specs.yaml
[value](#=)      pending — not yet verified

Tables and figures: mark the caption, not individual cells.

[**Table 1: Results by model.**](#=table1)

| Model | Accuracy |
|-------|----------|
| GPT-4 | 84.7%    |

Specs in specs.yaml

Two reserved fields, everything else is a template variable:

clf_lr_acc:
  prompt: "Run: python eval/classify.py. Report logistic_regression.accuracy from the JSON output."
  note: "5-fold stratified CV, seed 42."

Use YAML anchors to DRY up repeated patterns:

_templates:
  classify: &clf
    prompt: "Run: python eval/classify.py. Report {model}.{metric} from the JSON output."

clf_lr_acc:
  <<: *clf
  model: logistic_regression
  metric: accuracy

clf_rf_acc:
  <<: *clf
  model: random_forest
  metric: accuracy

{var} references are expanded at build time. Only prompt and note reach the reader.

Project config (livepaper.yaml)

title: "Your Paper Title"
setup: |
  Help me navigate this replicable livepaper. Clone https://github.com/org/repo and cd into it.
  Run: pip install -r requirements.txt
  You are now ready to replicate any result.

The setup field is shown in a banner at the top of the built page with a copy button.

Repo structure

your-paper-repo/
├── paper/                    # your original LaTeX (untouched)
│   └── paper.tex
├── livepaper/                # created by livepaper init
│   ├── INSTRUCTION.md        # agent playbook
│   ├── paper.md              # markdown + markers
│   ├── specs.yaml            # replication prompts
│   ├── livepaper.yaml        # project config
│   ├── assets/figures/       # figure images
│   └── dist/                 # built output
│       ├── index.html
│       └── specs.json
├── ...                       # other necessary files

License

MIT