Flownix
Начать
Sections
On this page

A new product from scratch

An intent and an empty repository. What it takes to get a tree of work you can start from.

The situation

You are starting a new product. You have an idea of what it should become and nothing else: no code, no tasks, no documentation. You will be working with an AI agent.

The usual alternative is a board and a list of tasks. A month later the list has grown, some tasks are duplicates, and nobody remembers why half of them were filed.

What you do

State the intent, not the tasks. A sentence or two about what you want to achieve. Decomposition is the agent's job, and doing it for them throws away their strength.

Check the decomposition. Once the tree exists you look at it and cut: what is not needed this quarter, what is too coarse, which dependency is spurious. This half of the work is entirely yours.

Say what the agent cannot know. Deadlines, priorities, agreements.

What the agent does

in the agent session, in order
/flownix-init
/plan-feature A meeting-room booking service: an employee sees free slots
for the week ahead and books one; a booking is cancelled by its author or an admin.

What sits behind that:

  1. list_organizations → create_organization, list_projects → create_project, then .flownix next to the code. This is also where you name the project key that slugs are built from: key MR gives MR-TASK-1.
  2. rag_query on the meaning of the intent — empty on a fresh project, but the check is always made.
  3. create_node top-down: feature → plan → task, with the shallowest tree that fits. An epic only for a genuinely large body of work.
  4. create_doc_node — the spec: what the product becomes, not what must be done. Tasks are linked to it with set_references.
  5. add_dependency for what blocks what, add_tag for tags.

The agent finishes by naming the slugs it created: MR-FEAT-1, MR-PLAN-1, MR-TASK-1…5, MR-DOC-1. From here on you work with them by name.

What comes out

A project where what to do, in what order and why is visible. The tree in the web app, the dependency graph beside it, the product spec as its own document.

From then on any agent — not only the one that planned it — can pick up a task and execute it, because the task is self-contained: it carries context and acceptance criteria.

Limits

The agent will not invent the product for you. It decomposes an intent; it does not generate one. "Build me a startup" yields a tree of generalities.

The first decomposition is almost always excessive. Agents lean towards thoroughness, and thorough is not the same as needed. Expect to cut.

An empty project knows nothing. Its semantic search is empty and fills up as you work. The first tasks are planned without the support of the past, because there is no past.

Do not turn on strict mode immediately. On a project without documents it hits the gate on the very first task. Work for a while first.

Next