Audit trail

Keep an Auditable Claude Research Notebook

A Claude research notebook is the bridge between a helpful chat session and a defensible research artifact. It records what was searched, what Claude saw, what changed, and what a human verified.

Every output carries an auditable history.

Separate the notebook from the chat

Chat history is convenient but not a durable research record. A notebook should live in a place your lab already trusts: a versioned repository, electronic lab notebook, secure shared drive, Zotero notes, or a project-management system with exports.

For every meaningful AI-assisted step, record the date, product surface, prompt, input files or source set, output, and human decision. If the step produced a figure, table, or code file, save the file and the generation path together.

Track source state as a first-class artifact

Research often fails because source state is implicit. The notebook should include search strings, database names, result counts, exported metadata, included/excluded record IDs, and the date of each search. For living reviews or fast-moving fields, add a planned refresh cadence.

When Claude is used with connectors, MCP, or Claude Science, document which external systems were reachable. This is especially important for proprietary datasets, lab pipelines, and internal notebooks.

Preserve correction history

The most valuable AI notebook entries are not the ones where Claude was right. They are the corrections: a hallucinated citation, a misread figure, a missed exclusion criterion, or a synthesis paragraph that overstated causality. Record the correction and the prompt or guardrail you added afterward.

Over time, this becomes a local methods asset. Your lab learns which prompts, source formats, and verification checks work for its field.

Make the final artifact cite the notebook

A manuscript, internal memo, or grant background section should be able to point backward to its evidence table and notebook. That does not mean publishing every prompt. It means preserving enough that an internal reviewer can reproduce the path from source to sentence.

For public-facing work, disclose the AI-assisted steps at the level required by the venue, journal, funder, or institution.

Workflow checklist

  1. Create a project notebook before using Claude.
  2. Record prompts, model/product surface, date, and input set.
  3. Save database queries and exported records.
  4. Store evidence tables and source verification notes.
  5. Log corrections and changed instructions.
  6. Attach generated code, figures, and environment details.
  7. Reference notebook artifacts from the final memo or manuscript.

Researcher FAQ

Do I need to save every Claude message?

Save every message that materially affects the research result. Casual brainstorming can be summarized, but source selection, extraction, analysis, and drafting steps should be preserved.

What belongs in a public disclosure?

Follow the journal, funder, or institutional policy. At minimum, record the AI task, human verification process, and whether the tool contributed to search, extraction, analysis, or writing.

Related workflows