Simple explainers from long-form academic publications

Turn research PDFs into structured JSON explainer articles using the Claude or OpenAI batch/sync APIs.

• Input: local PDFs and/or a list of paper URLs.
• Output: one output/<slug>.json per paper — a validated, schema-stable explainer object — plus an optional self-contained output/<slug>.html.
• Cost: uses the Claude Message Batches API / OpenAI Batch API (50% discount) or a synchronous path for fast single runs.
• Determinism: model schema drift is normalised after generation; figures are extracted from the source PDF deterministically, never hallucinated.

What this is / is not

• It is a focused pipeline that produces a specific JSON schema (src/types/explainer-json.ts) intended for a separate React renderer.
• It is not a general-purpose PDF summariser or chat tool.
• Figure extraction is macOS-first: it shells out to sips. On Linux the pipeline still runs but image blocks are dropped (see Troubleshooting).
• The website integration is optional. Without it you still get JSON.

Architecture

flowchart TD
  subgraph IN["Inputs"]
    A1["Local PDFs<br/>(input/*.pdf)"]
    A2["Paper URLs<br/>(input/urls.txt)"]
  end

  IN --> B["Preprocess"]

  B --> C{"Run mode"}
  C -->|"batch (50% discount)"| D["Submit + poll provider batch"]
  C -->|"--sync (fast, single)"| E["Synchronous run"]

  D --> F["Validate + one-pass repair"]
  E --> F
  F --> G["Normalise schema drift"]
  G --> H["Extract figure from source PDF"]
  H --> R(["output/&lt;slug&gt;.json"])

  R -. "only if WEBSITE_REPO set" .-> S["Stage + render website HTML"]

  classDef opt stroke-dasharray:4 4;
  class S opt;

Claude and OpenAI are both supported as the provider; the website step (S) is entirely optional.

Prerequisites

• Node 20+ (.nvmrc pins 20; package.json enforces engines).
• One provider credential — an Anthropic or OpenAI API key, a Claude Max/Pro OAuth token, or Codex CLI auth (see Auth routes).
• poppler — pdftotext, pdftoppm, pdfimages, pdfinfo, used for figure extraction. brew install poppler / apt-get install poppler-utils.
• sips — ships with macOS; used to downscale extracted figures. Not available on Linux: figures are skipped there, the rest of the pipeline is unaffected.
• tmux — only required for the secure headless route (npm run process:secure:tmux).

Quick start (no 1Password — primary path)

npm install
cp .env.example .env
# edit .env: set ANTHROPIC_API_KEY and/or OPENAI_API_KEY

# Provide input as a local PDF...
cp ~/Downloads/my-paper.pdf input/
# ...and/or paper URLs (arXiv, DOI, publisher page, or direct PDF link):
echo "https://arxiv.org/pdf/2411.13768" >> input/urls.txt

# Claude batch (50% discount, ~1h turnaround):
npm run process -- --provider claude

# or OpenAI batch:
npm run process -- --provider openai

# or a fast synchronous run (Claude OAuth or Codex CLI auth, no batch wait):
npm run process -- --provider openai --sync

.env is loaded fill-only (src/env.ts): a value there is used only if the variable is not already set, so it never overrides a real shell export or a 1Password-injected key.

Generated JSON lands in output/ (or EXPLAINER_OUTPUT_DIR). No website repo is required — staging and website-HTML export are skipped unless WEBSITE_REPO is set.

Quick start (with 1Password — optional)

For non-interactive/headless runs without putting keys in .env or your shell:

cp op-refs.local.sh.example op-refs.local.sh
# edit op-refs.local.sh with your real op:// vault refs

npm run process:secure:tmux                 # default provider: openai
PROVIDER=claude npm run process:secure:tmux # Claude batch

scripts/run-batch-tmux.sh sources op-refs.sh, resolves only the key the selected route needs via an op-fetch resolver on PATH, and runs the child in a detached tmux session with a sanitised environment. If op-fetch or op-refs.local.sh is absent it falls back to running directly with .env.

Commands

Command	What it does
npm run process	Full pipeline: preprocess → submit batch → poll → save JSON (+ optional HTML)
npm run submit	Preprocess + submit batch, then exit (prints batch ID)
npm run poll	Poll the latest pending batch and save results when ready
npm run poll -- <batch-id>	Poll a specific batch by ID
npm run status	Show all batches, per-request results, token counts
npm run process -- --sync	Run synchronously (no batch queue)
npm run export-html -- --input-dir output	Re-render HTML from existing JSON (needs WEBSITE_REPO)
npm run typecheck / npm run build	Type-check / compile to dist/
npm run smoke	Offline check: every provider tier resolves and has a pricing row (no API calls)
npm run guards:install	Arm the local publish git hooks (contributors)

Always pass --provider claude or --provider openai explicitly. With no flag and no PROVIDER env var the default is claude.

Inputs

• Local PDFs — drop any .pdf into input/. OpenAI uploads are cached in state.json by filename + content hash; unchanged files are reused.
• Remote URLs — input/urls.txt, one HTTP(S) link per line (# comments ignored).
• Per-paper focus hint (optional) — steer emphasis for one paper without changing the global prompt:
  • PDF: sidecar input/<basename>.focus.md (body = emphasis block).
  • URL: append # focus: … to the line in urls.txt.

Per-paper directives (<paper>.focus.md)

For a PDF at input/my-paper.pdf, an optional sidecar at input/my-paper.focus.md lets you steer one paper without touching the global prompt. Structured directive lines are stripped before the model sees the prose; the remainder is appended verbatim to the user message under the heading READER EMPHASIS FOR THIS PAPER. Malformed directive lines are ignored silently.

Directive	Purpose
image: Figure N	Force a specific figure as the lead image (accepts Figure 3, Fig. 4a, etc.)
image_caption: …	Override the caption attached to the lead image
image_alt: …	Override the alt text for the lead image
image_page_hint: N	1-based page number used by the figure cropper instead of caption search

See input/.focus.md.example for a copy-paste-and-edit template.

Configuration

All variables are optional except a provider credential. Set them in .env, .env.local, or the shell.

Variable	Purpose	Default
ANTHROPIC_API_KEY	Claude batch / API	—
OPENAI_API_KEY	OpenAI batch / API	—
CLAUDE_CODE_OAUTH_TOKEN	Claude Max/Pro sync route	—
PROVIDER	Default provider when no --provider flag	claude
MODEL_BATCH MODEL_LANE MODEL_SYNTHESIS MODEL_REPAIR	Per-stage model overrides	see below
MAX_TOKENS LANE_MAX_TOKENS SYNTHESIS_MAX_TOKENS REPAIR_MAX_TOKENS	Per-stage token caps	see below
EXPLAINER_INPUT_DIR	Where source PDFs / urls.txt / focus sidecars are read	<repo>/input
EXPLAINER_OUTPUT_DIR	Where JSON is written	<repo>/output
WEBSITE_REPO	Consuming website repo; enables staging + website HTML	unset (skipped)
EXPLAINER_JOBS_DIR	Shared batch-dashboard jobs dir (best-effort)	<repo>/jobs

Model defaults — Claude: batch/lane/synthesis claude-opus-4-8, repair claude-sonnet-4-6. OpenAI: batch/synthesis gpt-5.4, lane/repair gpt-5.4-mini. For a committed profile, copy config/models.json.example to config/models.json (gitignored; env vars still win over the file).

Auth routes

Route	Credential	Notes
Claude batch	ANTHROPIC_API_KEY	50% discount, ~1h turnaround
Claude sync	CLAUDE_CODE_OAUTH_TOKEN	Max/Pro quota; mixed OAuth+API-key env is rejected
OpenAI batch	OPENAI_API_KEY	semantic-lane extraction + synthesis
OpenAI sync	Codex CLI auth (no key fetched) or OPENAI_API_KEY	fast single runs, no batch wait

1Password is optional and orthogonal to the route — see Quick start (with 1Password) and docs/SECURITY.md.

How it works

1. Preprocess — read local PDFs and the optional URL list.
2. Submit — Claude: one explainer request per paper. OpenAI: semantic lane extraction (methods/results/limitations/implications) with a smaller lane model.
3. Poll — exponential backoff (30 s → 5 min cap) until the batch ends (skipped for --sync).
4. Synthesis + save — OpenAI runs a second synthesis stage; all providers get validation + an optional one-pass repair, then:
   • normalizeSchemaDrift derives canonical paragraphs and end_takeaway.label/body when the model emits *_html/heading variants, so the renderer never shows blank blocks.
   • attachFigureImage applies any lead-image override, then deterministically locates the named figure in the source PDF (poppler + sips), re-encodes to a downscaled JPEG, and inlines it as a base64 data URL. Failures drop the image block silently — the explainer still renders.
   • JSON is written to the output dir; if WEBSITE_REPO is set it is also staged and rendered to standalone HTML.

The system prompt is skill.md (after its YAML frontmatter) sent verbatim.

Website integration (optional)

Only relevant if you set WEBSITE_REPO to a checkout of the consuming website repo. When set, every saved JSON is copied into that repo's explainers-new/ staging dir, and html-export.ts spawns its scripts/export-explainer-html.mts to write a self-contained HTML file next to the JSON (Chart.js from CDN, no Next.js runtime dependency). output/*.html opens directly in a browser for a pixel-accurate preview.

Shared type contract: src/types/explainer-json.ts here must stay in sync with the website repo's equivalent ExplainerJson type. Any schema change must land in both repos.

Troubleshooting

Symptom	Cause / fix
Website HTML export skipped (WEBSITE_REPO not set)	Expected when not using the website integration. Not an error.
Image silently dropped	poppler not installed, or on Linux (sips is macOS-only), or the named figure was not found in the PDF.
JSON parse failed — raw output saved to *_error.txt	Model returned non-JSON; inspect the .txt in the output dir.
Claude run rejected for mixed auth	Both CLAUDE_CODE_OAUTH_TOKEN and ANTHROPIC_API_KEY were set for a sync run. Use the secure tmux route or unset one.
op-fetch is not installed	Not fatal — the secure wrapper falls back to .env. Only the 1Password route needs op-fetch.

Development

npm run typecheck      # tsc --noEmit
npm run build          # compile to dist/
npm run guards:install # arm local publish hooks (one-time, per clone)

state.json is auto-managed — do not hand-edit. Agent guidance lives in AGENTS.md; Claude Code specifics in CLAUDE.md. The private-work → public-mirror branching/publish model is documented in docs/PUBLISH-WORKFLOW.md.

Security

Credentials are never committed. The .env/.env.local loader is fill-only; the 1Password route resolves only the keys a route needs and runs the child with a sanitised environment; committed op-refs.sh holds placeholders while real refs live in a gitignored op-refs.local.sh. Local git hooks (scripts/install-guards.sh) block machine-specific data from being committed. See docs/SECURITY.md.

License

MIT — see LICENSE.