Turn notes into a document
Put a small set of notes in a permitted workspace. Ask for a report or slide deck with an outline, clear source attribution and a named output file. The documented office tools support Word, Excel and PowerPoint workflows and rendering for review.
Read the notes in this folder. Draft a two-page project update in DOCX, with decisions, open questions and next steps. Cite the source note for each decision. Save it in outputs and render it so we can check the layout.
Open the result and inspect page breaks, tables and missing material. Rendering and format conversion can change layout; review the actual artifact instead of assuming that a successful file write proves the document is correct.
Analyze a spreadsheet or CSV
Python and the bundled data tools can inspect tables, calculate summaries and produce charts. Start by asking the agent to describe the data and its assumptions before it computes totals. For a spreadsheet with formulas, check that the workbook opens and that results match the intended calculations.
Inspect orders.csv. Flag missing dates and duplicate order IDs. Summarize monthly revenue, keeping refunds separate, and make a chart. Save the cleaned data as a new file and list every cleaning rule.
Compare a few rows and totals manually. Ask about ambiguous dates, currencies or missing values. Use a copy of the source when experimenting.
Build or change code
Local Code includes Python, Node.js, a shell and git inside its sandbox. Give it a narrow change and ask it to inspect existing instructions, preserve unrelated work and run the relevant checks.
Read this project's README and tests. Add CSV export to the existing report page, preserving the current design. Run the relevant tests and show the changed files. Do not commit or publish anything.
Review the diff and exercise the resulting behavior. The in-app shell cannot launch arbitrary programs installed elsewhere on your Mac. Package compatibility also matters: a package that needs a native extension or host executable may not work like a pure Python or JavaScript package. Downloads need connectivity.
Research with sources
Web search and the browser tools can help gather current information, but they require network access. Local inference does not make web research offline. Specify the sources, date range and evidence you need.
Compare the current documentation for these three APIs. Use their official websites, link each claim to a source and separate confirmed behavior from your inference. Save a comparison table and note anything you could not verify.
Check links and dates yourself before relying on important conclusions. If you select a cloud model, relevant prompts and tool outputs may also be sent to that provider. See offline use and privacy for the separate model and tool boundaries.
Build a repeatable workflow
Once a small task works, save its requirements and review checklist with the project. The official repository documents built-in tools and skills, and custom skills can capture recurring instructions. Keep checking generated work; a repeatable prompt is not a guarantee of an identical or correct result.
New to the app? Start with Mac setup and choose a model that fits your hardware.