Keep-Asking Task Designer

A small kickoff helper for the keep-asking study (HREC 83897). A co-investigator uploads an existing worksheet (or pastes a rough idea), and the tool returns 2–3 concrete AI-assisted lab task designs in the study's approved format — scenario, brief, deliverable, a 1–5 rubric, the 2–4 load-bearing specifics a generic AI answer would miss, and an honest pre-pilot checklist. It can also stress-test a task you already have and refit it. You can then iterate ("make option 2 harder; add a conflicting stakeholder") and download each design as Markdown matching docs/task-design-template.md.

It exists because the hardest part of the kickoff ask is exactly this: designing a task where challenging the AI pays off. The tool is itself a little dialogue-over-delegation exercise — you converse with the model to sharpen the task.

Privacy model

• Files are parsed in the browser (mammoth for .docx, pdf.js for .pdf). The raw file never leaves your computer — only the extracted text is sent.
• The server is a stateless proxy: it holds the Anthropic API key, forwards the request, returns the result, and stores nothing (no database, no disk writes).
• Extracted text is sent only to the Anthropic API, with provider training opt-out recommended on your account.

Run locally

cp .env.example .env        # add your ANTHROPIC_API_KEY
npm install
npm start                   # http://localhost:8080

Deploy on a VPS (Docker)

Images are built and pushed to GHCR by GitHub Actions on every push to main (.github/workflows/docker-image.yml). On the VPS you only need docker-compose.yml and a .env file:

# one-time
cp .env.example .env        # fill in real values

# each release
docker compose pull
docker compose up -d

Put a reverse proxy (Caddy/nginx/Traefik) in front for HTTPS and access control; the container listens on :8080 and has no built-in auth. For a small team, a one-line basic_auth in your Caddyfile is the simplest gate:

keep-asking-task-designer.example.edu {
    basic_auth { team JDJhJDE0...hash }   # generate with: caddy hash-password
    reverse_proxy keep-asking-task-designer:8080
}

Configuration (.env)

Variable	Purpose	Default
ANTHROPIC_API_KEY	Required. Held server-side only.	—
MODEL	Quality generation model.	claude-sonnet-4-6
FAST_MODEL	"Fast/cheap mode" for quick iteration.	claude-haiku-4-5
MAX_REQUESTS_PER_DAY	Per-visitor-IP request cap.	100
DAILY_TOKEN_CAP	Global output-token ceiling/day (then 503).	2000000
MAX_INPUT_CHARS	Reject oversized source text.	60000
PORT	Listen port.	8080

Security notes

• No login in the app by design. Access control is delegated to the reverse proxy (e.g. Caddy basic_auth). The app adds per-IP and daily-token caps on top. This suits a small team over a short window — it is not hardened multi-tenant auth. Don't widely advertise the URL.
• The rate/spend guards are in-memory and reset on container restart.

What it is not

Generated task designs are drafts for the team to review and sign off before piloting. The tool encodes the study's design rules but does not replace investigator judgement or the ethics process.