Research kit

AI Research Starter

For anyone interested in AI research: a starter pack to help you and your AI assistant plan, investigate, challenge and document a project that others can check. This is a public pilot edition, shared for trial and feedback. It is not evidence that the pack improves research quality.

01Download

AI_Research_Starter.zip (222 kB, 47 files, edition 2026-10-16.1-public-pilot).

Unzip it, read README.md, then fill in CLAUDE_START.md and give its prompt to your assistant. The prompt also works with other assistants.

To check the download, compare its SHA-256 with ef3e374781c47980016b1578d79294c37a4488d482d702b5c9dbed9e54a2496f.

02What is inside

  • A starting prompt with project details to fill in (CLAUDE_START.md).
  • Short working rules, the research workflow, and a method menu with short route notes (WORKING_RULES.md, WORKFLOW.md, METHOD_SELECTION.md and methods/).
  • Guides for checking prior work, sources and reproducible calculation, analysis checks, an optional separate review and a fair note on AI help (IDEACHECK.md, SOURCE_AND_REPRODUCTION.md, ANALYSIS_CHECKS.md, CROSS_REVIEW.md, AI_DISCLOSURE.md).
  • A project notebook for scope, evidence, decisions and what was done (PROJECT_NOTEBOOK.md).
  • Two small checkers with their tests: prosecheck.py looks for configured house-style and internal-marker patterns, and does not detect facts or authorship; vibecode_gate.py runs bounded Python and JSON checks.
  • A worked example: an earlier, unchanged version of our AI infrastructure paper with its reproduction script (examples/ai-infrastructure-v7.1/).

03What you need

An AI assistant, and Python 3 for the checkers and the example; they use only the standard library. The pack's Windows instructions have not yet been tried on Windows.

04Licence

The documents are shared under CC BY 4.0, and the scripts (prosecheck.py, vibecode_gate.py and tests/) under the MIT licence. The example in examples/ keeps its own CC BY 4.0 licence. The full terms are in LICENSE.md inside the zip.

05Limits

Following the instructions or passing the scripts does not show that a source supports a claim, or that the research is correct. You own the substantive decisions and any release, and you follow any research-ethics, data-use or other rules that apply to your work. The pack does not replace advice from a subject specialist.

06Feedback

Observations from using the pack are welcome at NMAIResearch@proton.me. Please remove private work, personal data and credentials first. How email to that address is handled is on the privacy page.