Implementation prompt builder¶

Prepare instructions for Claude or another coding agent. Choose paper/code sources, a new local or GitHub repository, and automatic or explicit profile selection. Evaluation tasks, if supplied, must already have labels and task configs. This form assembles text locally; it does not read your files, contact GitHub, upload inputs or run an agent.

The generated prompt uses the same maintained instructions as the static workflow prompt. You can edit the result before copying. Demo and full-data execution retain their human review gates.

1. Sources Available source material

Paper references One or more URLs or local paths; include supplementary material and versions.

Code references Repository URLs or local paths, commit/tag, relevant configs and checkpoints.

2. New repository Destination New local repository path Also required for a GitHub destination: this is the agent's working checkout. Choose a directory that does not exist yet.

3. Profiles How should profiles be chosen?

Automatic selection happens in the implementation agent, before generation. It may select several profiles when the sources support them.

4. Dataset and prepared tasks Dataset context Name, revision and MEDS directory. Required when supplying tasks; otherwise optional. Evaluation task inputs

The prompt does not ask the agent to invent tasks or extract labels. Missing or inconsistent inputs defer only the dependent evaluation.

5. Scope and constraints Requested endpoint

Preparing for full evaluation does not authorize a demo, dataset download or cluster run. The existing review gates still apply.

Additional constraints (optional) For example: algorithm name, hardware limits, checkpoint requirements or implementation priorities.

Fill in the inputs, then generate your prompt.

Editable implementation prompt