Best Practices

Organize by Research Boundary

Preserve Raw Data

• Keep authoritative source files immutable and backed up outside CATO as required by your governance plan.
• Upload a data dictionary, protocol, analysis plan, and codebook alongside the data they define.
• Store cleaned and analysis-ready datasets as new artifacts; do not silently replace the source.
• Record units, missing-value codes, reference build, time zone, locale, and categorical reference levels.
• Use stable keys to validate joins and check the expected row count after every merge.

Select Context Deliberately

Mention exact datasets with @, select the intended artifact version, and attach only the literature needed for the question. If two files have similar names, state which is authoritative and how they differ.

Plan Before Consequential Work

Press Shift+Tab for Plan mode when work includes exclusions, imputation, model selection, multiple endpoints, evidence eligibility, external lookups, expensive execution, or publication output. Review the plan for hidden assumptions and ask for revisions before approving execution.

A good plan names inputs and versions, transformations, method, diagnostics, expected artifacts, stopping conditions, and what CATO should ask you rather than infer.

Choose the Right Mode

• Fast: rapid exploration, summarization, formatting, and iterative edits.
• Standard: routine analysis and writing where balance matters.
• Research: complex methodology, deeper evidence work, and extended tool use.

Mode is not a substitute for a precise prompt. Start with the least intensive mode appropriate to the task, then escalate when depth—not ambiguity—is the limiting factor.

Separate Conversations by Objective

Keep one coherent analytical thread per conversation. Start a new conversation when switching from exploration to a prespecified primary analysis, from one manuscript section to another, or from one evidence question to an unrelated one. This reduces context contamination and makes durable runs, reviewer findings, and artifacts easier to audit.

Handle Durable Runs Correctly

• If a response reconnects after navigation, wait for recovery before submitting a duplicate.
• Use background jobs for long execution and inspect the persistent job card after reload.
• Cancel once when the inputs or objective are wrong; wait for the terminal state.
• After an interruption, check for finalized text and artifacts before retrying.
• Name retries clearly so their outputs are not confused with the original run.

Use the Workspace as the Record

Open output files, notebooks, figures, and manuscripts in the Workspace. Review provenance and logs, select the exact version used downstream, and save or discard unsaved edits before switching versions. Chat summaries are useful navigation; artifacts and their source evidence are the auditable record.

Make Notebooks Re-runnable

• Put imports, configuration, seeds, and data loading in explicit early cells.
• Avoid hidden dependence on cell execution order or manually edited in-memory objects.
• Save important intermediate data and parameters as artifacts.
• Restart the persistent kernel and run from a clean state before finalizing.
• Export code and environment details required for an independent reproduction.

Treat Figures as Data Products

• Preserve the plotting data, code, and original figure version.
• Use figure versions to distinguish analytical revisions from visual edits.
• Check axes, transformations, denominators, error bars, legends, clipping, and annotations.
• Use accessible palettes and verify the figure at its final publication size.
• Run visual QA after editing and inspect the actual exported PNG/SVG/PDF—not only the browser preview.

Curate Literature Reproducibly

• Save the question, date, translated query, source selection, filters, and connector warnings.
• Use Novelty-first for discovery and Evidence-first for established evidence, then inspect the actual ranking.
• Resolve preprint/journal duplicates, corrections, and retractions.
• Verify evidence extractions and study eligibility against full text where available.
• Use watches to stay current, but account for indexing and source delays.
• When sending papers to Chat, bound the context and require source support for each claim.

Use Memory for Preferences, Not Findings

Project memory can hold stable protocol language or definitions supplied by you; global memory can hold durable writing or visualization preferences. Review, edit, pin, or delete entries and control recall, automatic learning, and cross-project behavior.

Validate Before Sharing

Share Safely

• Prefer an exported, reviewed artifact over screenshots that omit context and provenance.
• Before creating a public conversation or figure link, confirm that no restricted data, prompts, outputs, or metadata are exposed.
• Use the narrowest sharing mechanism and revoke public links when no longer needed.
• Do not place credentials, secrets, protected health information, or confidential raw data in support tickets.
• Follow institutional data-use, authorship, disclosure, and retention requirements.

Keep an External Research Record

CATO provides durable production storage and versioning, but it is not your entire records-management policy. Export final data, code, notebooks, environments, literature searches, evidence tables, manuscripts, figures, and audit materials to the approved repository required by your institution, protocol, funder, or regulator.