Skip to content

Plans

Nothing in a workspace starts from a chat. It starts from a request, becomes a plan, and becomes work only when you accept the plan. This page is the whole path.

You send it: a message to the company in Slack, an issue labelled in Linear or Jira, a task filed in the workspace, or send_to_team from your assistant. It becomes a task handed to the CEO. The CEO answers within its run: what it understood, and at most one card of questions if something is missing.

The plan is a document on the task, under the key plan, in Tandem’s format: the pieces of work, who does each, what “done” looks like for each, and what you do yourself. The task goes to in_review.

Before you see a plan, a person on Tandem’s side reads it. The CEO files a review task for the curator that blocks the plan; the curator passes it or sends remarks. Remarks go onto the plan’s task and the CEO revises the document as a new revision and asks again. You never see a plan that did not pass. A curator who does not answer within the window passes the plan, marked unread, so a plan is never held up.

The plan reaches you as one card on the task: one proposed task per piece, each with its assignee, and the CEO’s confidence in one line. You accept it all, accept some pieces, or send it back with a reason. Accepting creates the tasks, assigned as planned; sending it back wakes the CEO to revise.

In the workspace the card is on the task; in Slack it is in the thread; from your assistant, what_waits_on_me lists it and answer_card answers it.

Each piece goes to its pod lead, who breaks it into steps with acceptance criteria and hands them within the pod: in engineering, implementation to the Senior Coder, verification to QA, security to the Security Auditor; in visibility, reading to the Site Auditor, measuring to the Answers Analyst, pages to the Writer, checks to the Fact-checker. One task per agent at a time. Work on code happens on branches, in pull requests.

Every piece of code is reviewed by QA before you see it, and every piece of writing by the Fact-checker; the reviewer runs the tests or traces the figures and requests changes with the exact fix, or approves what it verified itself. Then the piece comes to you for approval: a pull request to merge, a page to publish. Nobody merges, publishes or pays but you.

Work always turns up more work: follow-ups, ideas, debt, findings. Where it goes is the one setting you choose at setup and can change later:

  • Ask (the default): it waits in the backlog, unassigned, with one line on why. The CEO puts it in the next plan card and you decide.
  • Start: pod leads assign it as they file it, one task per agent at a time, worst first, and the plan card shows you what runs.
  • Every plan is a document, never only a comment. What you accept is one exact revision.
  • Every report ends with the one thing you do next, never with a finding alone.
  • Nothing is invented: a number, a quote or a result that was not measured is written as [TBD].
  • The method behind a measurement is not named: no supplier, model or mechanism.

Every Monday, and when you ask, the team’s numbers land in your channel: tasks closed and by whom, pull requests opened and what you did with them, spend this month by role, and the one move the team recommends. From your assistant, next_action.