CATO Chat and Workspace

Analysis Modes

Change modes with the model control in the composer; /standard switches back to Standard. The mode is a product behavior contract; CATO may route work across specialized analysis, reading, research, or coding capabilities as the task requires.

Fast mode is retired, since Standard now runs at Fast speed. Chats that used Fast run on Standard.

A separate reasoning-effort slider in the composer sets how much CATO thinks before answering: Low for fast replies to simple questions and small edits, Medium as a good default for most analyses, High for extended reasoning on every step, and Max for full reasoning at the cost of speed and allowance. The effort applies to the message you send and the run it starts, and your last choice becomes the default for new messages.

Build the Context

CATO performs best when the active evidence is explicit. Use @ to mention project datasets, select workspace artifacts or versions, or send a curated literature set from the Library. Context cards let you see what will accompany the prompt.

• Refer to exact filenames, versions, variables, cohorts, and timepoints.
• State whether CATO may combine context with general knowledge or must use only selected sources.
• Remove irrelevant attachments; more context is not always better context.
• For follow-up work, confirm that the intended generated dataset or figure version is selected.

Compose and Control a Prompt

Ask CATO to plan first and it presents a reviewable approach before acting. Do this when the method, inclusion criteria, transformations, model specification, or expected outputs need your approval. You can also ask CATO to revise the plan rather than approving it as written.

A Reproducible Prompt Pattern

Following a Run

While CATO works, the status line shows what it is doing, including the tools it is using, and long thinking shows a timer rather than going quiet. For multi-step autonomous work, a run rail appears once CATO has declared a plan and tracks progress through it. CATO may also pause to ask you a clarifying question mid-task when the request is ambiguous; answer it in place to continue.

• Claims in the response carry quiet marks tying them to the evidence that backs them; unbacked claims are flagged for your attention.
• A context meter shows how much of the conversation's working context is in use.
• Mathematical notation renders as typeset math, and prose renders as standard formatted text.

The Project Computer

Every project has its own durable computer where analyses run and files live. The Workspace panel names it in its title bar, with pinned Files, Shell, and Memory views. Files created by an analysis persist across conversations, uploads land in the folder you are browsing, and everything shows where it came from. You can rename the computer, and a cold start can take up to a minute before the first command runs.

Durable Responses and Reconnection

Active responses are conversation-scoped and persisted outside a single browser request. Navigation, refreshes, frontend restarts, and transient connections do not necessarily discard the run. Reopening the conversation reconnects to available stream events and finalized output.

• A reconnecting indicator means CATO is recovering state; avoid immediately duplicating the request.
• Cancellation is durable: use Cancel once and wait for the run to acknowledge it.
• Finalization prevents a completed response from remaining indefinitely in an active state.
• Partial output may be preserved when a budget, provider, execution, or validation error interrupts the run.

Background Sandbox Jobs

Long code or scientific-computing tasks can continue as durable background jobs. Their cards remain associated with the conversation after reload and expose current state, progress messages, completion, produced artifacts, failure details, and cancellation.

The Artifact Workspace

Artifact tabs can be restored when you return to the conversation. The selected file version is part of the working state; unsaved-edit prompts protect local changes before a tab, version, or view changes.

Figures and Scientific Viewers

Generated figures appear inline and in the Workspace. Supported SVG figures can be opened in the editor to select, move, group, layer, restyle, and annotate elements. Figure workbenches preserve interactive context for supported scientific visualizations, including molecular and other domain-specific views.

• Save edits as a new version and retain the original; the version badge on a figure switches between versions, and the choice follows into the Citations page and downloads.
• Use accessible colors, legible labels, units, uncertainty, and complete captions.
• Run visual QA after edits; confirm that data marks and axes were not accidentally altered.
• Export only after checking raster/vector fidelity and the target journal's dimensions.
• A public figure citation link can identify a shared figure; do not use it for restricted data.

Notebooks, Code, and Execution

CATO can create Python or R code, run it in an isolated sandbox, and preserve notebooks and files as artifacts. Code runs in CATO's session, not yours: ask for the change you want and it re-runs the analysis. Durable jobs cover tasks that should survive a browser navigation.

• Inspect generated code, dependency versions, seeds, inputs, and output paths.
• Do not assume a package is present merely because a language supports it; ask CATO to check the environment.
• Sandbox networking, compute, storage, and job duration are governed by production security and capacity controls.
• Re-run critical results from a clean state or exported environment before publication.

Manuscripts, Readers, and Annotations

Manuscript artifacts support Markdown or LaTeX editing, preview, citation-aware review, and export. PDF, HTML, and transcript readers can retain anchored annotations where the format supports them. Reviewer findings can point to evidence and artifacts without overwriting your source.

Memory

CATO can recall user-provided preferences or stable facts from project memory or, when enabled, global memory. Memory entries are visible and manageable: edit, pin, delete, disable recall, disable automatic learning, or prevent cross-project use from the memory controls. Each project also keeps an editable memory file, open from the Memory view on the project computer, that CATO reads at the start of every turn and maintains as it learns your project.

• Project memory is appropriate for protocol choices, defined terminology, or stable study context.
• Global memory is appropriate for durable personal preferences such as writing or plotting conventions.
• Automatic learning is grounded in what you say; generated analysis findings do not become evidence by repetition.
• Review memory when CATO appears to carry an outdated preference into a new task.

Literature-to-Chat

Select works in Literature Library and send a bounded context packet to Chat. Prompt templates support asking with context, structured extraction, drafting, and planning next steps. Fetched full-text papers, when legally available, can also open in the Workspace for reading and annotation.

The reviewer can use prior evidence attached to the conversation, but you should still open the cited source and confirm that the claim is supported. Learn more in the Literature Intelligence guide.

Peer Review and Assumptions

Review controls can ask CATO to challenge methodology, trace claims to evidence, inspect artifacts, or surface assumptions. Treat the result as a structured second pass, not independent peer review. Resolve material findings in the underlying data, code, figure, or manuscript and preserve the revision.

Usage and Billing Interruptions

AI activity is metered against included credit or pay-as-you-go usage. CATO reserves usage before concurrent requests, updates the meter as work proceeds, and can stop an active operation if an account budget, rate, or concurrency boundary is reached. Settings → Billing shows the current meter, included balance, and upcoming metered charge where applicable.

Prompt Checklist

• Is the research question falsifiable and the population explicit?
• Are the correct files and versions selected?
• Are variable definitions, units, missing-value codes, and exclusions clear?
• Did you request diagnostics, uncertainty, sensitivity checks, and limitations?
• Did you specify which outputs should be saved as artifacts?
• Will you validate code, results, figures, and citations before use?

For applied examples, continue to Best Practices or use the Quick Reference.