A folder can contain everything needed for a decision and still be difficult to use. The problem is often the handoff: a useful note disappears between research, comparison and presentation. An AI answer that sounds complete does not tell a colleague which source supports its recommendation.
Orkas is an open-source, local-first desktop AI workspace. Commander can coordinate specialist agents around one project, while project instructions, selected Library sources and working files provide reusable context. Here is a small, inspectable way to use that structure for a decision brief.
This is an illustrative workflow, not a recorded product test. The inputs and output below are fictional. The article was prepared with AI assistance for the Orkas inclusion campaign; it is not an independent customer review.
Define the decision before choosing agents
Imagine a volunteer group choosing between two meeting-recording arrangements. Its decision is whether to keep typed notes or make recordings. Give the project three short source notes:
- A — Policy: Participants must explicitly agree before a meeting is recorded. If someone declines, use typed notes. Do not upload identifiable participant information to external services.
- B — Pilot: In two fictional meetings, typed notes captured the final decisions but missed the owner of one follow-up action. No recording was made.
- C — Requirement: Every decision brief must identify the recommendation, supporting sources, unresolved questions and the person responsible for the next check.
These statements are the entire evidence base. They do not establish that recording is better, cheaper or legally appropriate. A useful workflow should preserve that boundary.
Give Commander a deliverable contract
Keep the notes together in a project and select the relevant Library material. Ask Commander to organize the comparison and prepare a brief, using document specialists where available. For example:
Use only notes A, B and C. Prepare a one-page decision brief about our next meeting. Separate observed information from proposals. Cite a note beside every factual claim. Do not infer consent, invent an owner or make external uploads. Include a missing-information list. Return the brief for review before any sharing or operational change.
Commander provides a place to coordinate the stages. OfficeWorker can help turn the checked material into an office document; if a presentation is needed afterward, PptMaker can work from the approved brief. Agent availability and configured tools determine which stages can actually run. There is no need to involve a research agent when outside research would violate the instruction to use only the supplied notes.
Know what an acceptable output looks like
A complete illustrative brief could read:
Decision: Keep typed notes for the next meeting and add an explicit action-owner section. Treat this as a proposed operating choice, not a result measured in the pilot.
Evidence: Typed notes captured final decisions in the two fictional pilot meetings, but one follow-up owner was missing [B]. Recording requires explicit participant agreement; a refusal requires typed notes [A].
Reasoning: The observed gap concerns action ownership. A written owner section addresses that gap without assuming consent to record. The pilot does not show whether recording would improve accuracy.
Before the meeting: Ask the organizer to name a note-taker and confirm an action-owner check at the end. The organizer and note-taker are not yet identified; do not invent names [C].
Open questions: Who will review action ownership? Where will the notes be stored? Are all participants comfortable with the proposed process? No recording or external upload is authorized by this brief [A].
That output is useful even if the reader never opens Orkas: it demonstrates how to turn a small source set into a recommendation without disguising assumptions as observations.
Inspect the handoff and the data path
Before exporting or sharing, compare the brief against each original note. Check that a document or slide version retains the citations, uncertainty and missing owner. Attractive formatting should follow the evidence review.
Orkas keeps working context and outputs local by default, but local-first does not mean every AI request stays offline. Relevant content can be sent to the selected model provider or authorized tools. Choose a supported own-provider configuration, a compatible local endpoint, or Orkas-managed access according to the task's constraints. The core desktop app is free; model usage and optional official services can cost extra.
The public desktop repository is MIT-licensed and includes the desktop setup instructions. Start with a non-sensitive example and an output you can inspect. The practical benefit comes from keeping evidence, instructions and the final artifact connected.