All guides

Guide · 7 min read

Organize AI work around projects, not browser tabs.

AI chats multiply fast. After 30 you can't find the conversation where you made a decision. This is the structure that survives that.

Create one workspace per active outcome

A workspace is something you're trying to finish: a product launch, a purchase decision, a trip, a client account. Not a workspace per question. A workspace per question is how you end up with 40 workspaces and no memory of any of them. Use folders when one project grows real streams of work.

Under the hoodHow it works
  • A workspace is the durable project boundary for chats, notes, files, and ordering.
  • Manual ordering should persist and folder moves should not cross section boundaries accidentally.

Use chats to explore and notes to decide

Chats are for branching, follow-ups, and alternatives. Notes are for the plan, the constraints, and the next action. Move the decision into a note and link it back to the chat. That's what keeps a project legible after the conversation has grown to 200 messages.

Under the hoodHow it works
  • Chats preserve exploration and source context; notes hold the current durable state.
  • Record identity must remain separate even when both types appear in the same workspace.

Attach the source, not every nearby file

A model only needs the material relevant to the question. Search by exact name, attach the smallest useful set. It's faster, cheaper, and safer. For code projects, ignored-folder settings matter: node_modules will bury your actual source files.

Under the hoodHow it works
  • Exact-name ranking and ignored folders reduce accidental attachment of nearby or generated files.
  • The selected attachment set should be visible before it is sent to a model.

Make switching safe and natural

A loading skeleton beats one frame from another conversation. Same after deleting or reordering: what you see should update immediately and stay that way when you reopen.

Under the hoodHow it works
  • Cancel obsolete loads and commit data only when the selected record identity still matches.
  • Deletion and reordering should update presentation state immediately and persisted state on reopen.

Publish only the outcome worth keeping public

A public record should stand on its own: title, summary, useful images, a date. Strip the private setup. Use public discovery only when you actually want search traffic; otherwise a link-only record with an expiration does the job.

Under the hoodHow it works
  • Generate the public record from an explicit selection rather than exposing its entire source workspace.
  • Preview media, chronology, RSS, profile inclusion, and link-only state should derive from published metadata.

Sources and scope

We wrote this from Axiom's current code, regression tests, and recorded product checks. If a behavior has only passed a source test, we do not describe it as proven on every Mac. There are no invented benchmark numbers here.