Think it through.
Get a plan an agent can execute.
Talk the idea out like a whiteboard. Cut what doesn't belong. Then turn what's left into modules, tasks, and acceptance criteria your agent can pick up without asking questions.
scope engineTalk it out before anything touches code.
Architect Chat is the whiteboard. Float the idea, cut what doesn't belong, and see the real surface area before anyone writes code. Every decision lands in the plan as a change you preview and apply.
- Decide what's in the MVP
- Pressure-test a feature before you build it
- Find the edge cases now, not mid-build
- Split the work into tasks one agent can finish
A plan an agent can actually execute.
Not a chat log. Not a markdown file in the repo. What comes out is modules, tasks, and acceptance criteria, sized for one agent session, with an order to the work. Paste a brief, a ticket, or a one-liner, or talk it out first.
- Modules, features, tasks
- Acceptance criteria on every task
- Work in dependency order
- Estimates in your unit
Grounded in the repo you actually have.
Import the codebase and the engine reads it before it plans: your stack, your conventions, what already exists. Pin those defaults once and every new scope lands in a context your agent already understands. Nothing is written into the repo. The plan lives here.
- Your stack
- Your target agent
- How you like tasks written
- Import from GitHub
Then you watch the build as agents work.Track it →