Module guide
Smart capture
Turn a freeform sentence into a task and subtasks, no LLM required.
Overview
Type anything, such as "cook butter chicken tomorrow with subtasks for ingredients", "launch the new landing page", or "plan a meeting next week", and Atlas produces a structured draft with subtasks, due dates, priority, and labels. It is 100% deterministic: recipe and archetype dictionaries plus an AST parser, with no external API.
Highlights
The capabilities worth knowing before you dive in.
- Recipe dictionary: butter chicken, biryani, pasta, pizza, salad, stir fry, omelette, sandwich, and chicken curry
- Archetype dictionary: launch, meeting, travel, onboarding, event, and interview
- Inline subtask extraction from "with subtasks: A, B, C" phrasing
- Strips imperative framing ("add task of", "remind me to", "please")
- A hints panel explains every decision so the output is auditable
Important to know
Limits, permissions, and sharp edges to keep in mind.
- Explicit lists always win over dictionaries. If you say "with subtasks: A, B, C", you get A, B, C and nothing else.
- The same input always produces the same output, which is great for testing and trust.
How to use it
The primary workflow, start to finish.
- Click "AI capture" on any task composer (it is exposed on My Work, Capture, and the composer drawer).
- Type a freeform sentence, then hit Capture.
- Review the subtask preview and hint chips, edit any fields inline, then save the task. Subtasks are created under the parent automatically.
FAQ
- Where does the recipe knowledge live?
- In packages/shared-types/src/smart-capture.ts, where RECIPE_DICTIONARY is plain data. Add to it freely; tests and the UI pick up new entries on the next build.
- How do I stop Capture guessing subtasks I did not want?
- State them explicitly. An explicit list always wins over the dictionaries, so writing the subtasks you want produces exactly those and nothing inferred.
- Does this send my text to an external service?
- No. Capture is entirely deterministic and local to Atlas: dictionaries and a parser, with no external call. The same input always produces the same draft.
Automate this module
Everything on this screen is scriptable. Drive it from the REST API, or let an AI agent run it through the MCP server.
Was this page helpful?