Nebula Agent

Nebula Agent works beside your documents. It can answer questions, inspect project context, propose edits, and create deliverables while reporting back in the conversation.

Start with a clear outcome#

Open the agent with Cmd/Ctrl + L. State the result you want, the source of truth, and anything that must not change.

Turn the selected notes into a one-page decision brief. Keep the dates and quoted customer language unchanged. Draft an outline first and wait for my approval before editing.

A useful request usually names four things:

1. Source — the active file, selection, attachment, or @mentioned file.
2. Output — an answer, edit, chart, or new file.
3. Audience — who will read or use the result.
4. Constraints — facts, formatting, paths, or files to preserve.

Add the right context#

The active document is available automatically. Selected text can appear as a visible context item, which is useful for a targeted rewrite. Add supporting material with @, the attachment button, or a folder reference.

Use @research/market-notes.pdf and the Summary sheet in @data/forecast.xlsx. Update only the Evidence section in this document and cite which source supports each number.

Attachments can include supported text, PDF, and Office files. Remove irrelevant context before a narrow task; more context is not always better context.

Follow the conversation#

The agent talks back in the right panel. It may ask a question, explain a decision, propose a plan, or summarize the files it changed. Longer work can show todos and tool activity between those responses.

Expand

Treat activity as a progress record, not the deliverable itself. Open every modified file and inspect the actual document, workbook, or deck.

For a new subject, start a new chat so old context does not steer the next request. Conversation controls let you search prior chats and delete a conversation you no longer need.

Use planning for consequential work#

Planning can be enabled in Settings → AI. It is most useful when a task spans files or creates a deliverable you cannot quickly recreate.

Read @launch-review.pptx and @pilot-results.xlsx. Propose a plan for a three-page decision memo, including the source for each section. Do not create or edit files until I approve the plan.

For a small edit, ask directly. A one-sentence rewrite should not need a long plan or a crowded activity log.

Ask questions before edits#

The agent can analyze without changing files. Use that separation when the source is unfamiliar or high stakes.

Read this contract and list every renewal date, notice period, and payment obligation. Do not edit the document.

Then scope the change:

Add a five-row obligations table after the Executive Summary. Preserve the existing styles and do not change any contract language.

Work across formats#

The agent can combine project sources and create or edit supported documents. Typical requests include:

• Compare a DOCX brief with an XLSX source of truth.
• Turn a PPTX working deck into a durable decision memo.
• Extract claims from a PDF and map them to a LaTeX outline.
• Analyze CSV data, create a chart, and write the result into a report.
• Audit a presentation for overflow, sparse slides, or inconsistent formatting.

Use a new filename for generated output:

Create exports/board-update.pptx from @quarterly-review.docx. Keep the source document unchanged and report every file you create.

Review the result#

Before accepting an edit:

• Compare facts against the named source.
• Inspect page, slide, table, and chart layout.
• Check formulas and calculated values after workbook changes.
• Verify citations and links yourself.
• Use Discard All when the operation was based on the wrong scope.

Review changes & recover explains the review and history surfaces for each file type.

Models, reasoning, and usage#

Choose a managed model and reasoning level in the composer. Availability depends on the account and can change; the in-app picker is the source of truth. Higher reasoning can help with multi-file synthesis or complex structural work, while low reasoning is usually enough for direct questions and small edits.

Nebula Writer does not currently expose user-managed provider API keys. Account usage and model access are managed through Nebula Writer.