Discuss one product area

The Loft product

You ask, Loft builds.

Say what the product should do. Loft writes the specification with you and builds against what you approve, through your coding agent. Yard shows you every question, change and plan along the way, and the specification and the record of each change stay in your repository.

See it work

One change, in five stages.

A worked example on a fictional subscription service: pause and resume, from the authored contract to the human decision. Beside each stage, the change's Yard Log: what the agent asked, what you approved, and the work as it ran, committed with the branch for whoever reviews it.

  1. Author

    Requirements and acceptance criteria for pause and resume, with the decision behind them.

    Yard LogQuestionanswered

    Can an account with an overdue charge pause its subscription?

    No: the request is refused until the charge is paid.

    RQ-9 If the account is overdue, then the service shall reject the pause request.

    • AC-36 A pause request from an overdue account returns account_overdue.
    • AC-37 A pause request made after the overdue charge is paid returns scheduled.
    • AC-38 A pause request from an account whose oldest unpaid charge is 13 days old returns scheduled.
  2. Ratify

    The reviewed change between two published states, line by line.

    Yard LogApprovalapproved

    The specification change for pause and resume

    Approved by the product owner, with the change drawn on the page.

    Changed RQ-3

    Before An active subscription is one whose status is active.

    After An active subscription is one whose status is active. A paused subscription is not active.

  3. Implement

    The pinned state and the affected requirements the coding agent works from.

    Yard LogPlanapproved

    Build pause and resume in four steps, each naming the criteria it meets

    Live: rejecting a pause from an overdue account, step 2 of 4

    Governing publish sequence 4

    loft-website-example

    RQ-3RQ-8RQ-9RQ-10RQ-11RQ-12RQ-13RQ-14RQ-15RQ-16

  4. Reconcile

    The first candidate's tests passed, and reconciliation still found a failure.

    Yard LogLoft reportkept

    Trace: one criterion failed, one has no test

    Kept as it was recorded, for the review.

    Candidate 1 · a7244b1

    Failed AC-36 A pause request from an overdue account returns account_overdue.

  5. Decide

    Reopened, then accepted, each on its evidence.

    Yard LogDecisionrecorded

    Reopened: an overdue account can still pause

    The fix ran as its own task, and the second candidate was accepted.

    1. Reopened Candidate 1 · a7244b1
    2. Accepted Candidate 2 · 11ec255

You decide, and it stays on the record.

People designated by the customer decide and ratify product intent. Loft records the accepted state, proposed changes, decisions and linked evidence in the specification, and Yard records how each change was made: what the agent asked, what was answered, and what was approved.

Work from the same contract across roles.

A private agent loop can make one specialist more productive. The delivery system still depends on what happens when work moves to the next specialist. Loft preserves the same product identities, decisions, requirements, constraints, and acceptance criteria across those transitions.

Product and domain work

Specialists develop and ratify outcomes, behavior, business rules, vocabulary, and priorities, and answer the agent's questions on the Yard page.

Architecture and engineering work

Specialists add technical constraints and decisions, approve the agent's build plan, then use the approved scope for implementation.

Testing and assurance work

Specialists add independent scenarios and evidence, preserving failures and unresolved claims for review.

Human acceptance

Authorized people review the connected intent, the result, the evidence and the record of how the work was done, rather than relying only on the latest handoff artifact.

More precise than a document, more durable than an agent transcript.

Loft can connect capabilities, requirements, components, constraints, quality targets, architecture decisions, vocabulary, and acceptance criteria inside the contract. It takes a specification-driven development approach: coding agents implement from one approved version of the contract, and reconciliation checks the result and its evidence against that same version, so reviewers judge it against what people actually approved.

The contract evolves through review.

A team can start with one bounded area and expand it over time. Each change to the specification comes to the people who own it on the Yard page before it is committed: they approve it, or send it back with what to change. Open questions and known gaps stay in view.

What waits on you.

  1. Answer

    Your agent asks rather than guesses, with a recommended option and its reasoning. Answer on the Yard page or in the agent's own dialog.

  2. Approve the change

    Each change to the specification is drawn on the page before it is committed. Approve it, or send it back with a reason.

  3. Approve the plan

    The agent plans the build from the criteria you approved, naming the criteria each step meets and the tests that will show it.

  4. Watch the build

    Tasks, subagents, loft runs and time, live, and the tokens each step takes where your agent reports them.

  5. Review

    Read the diff beside the branch's Yard Log, the agent's notes and Loft's reports.

  6. Decide

    Accept the work, send it back, or reopen it. Nothing is merged or published until you say so.

Change what has to be rebuilt at each handoff.

The next specialist and the final reviewer inspect the approved contract, the implementation revision, the mapped tests or measurements, the open gaps, and the branch's Yard Log of what was asked and decided, and judge the work on them.

The same authority model applies to new and existing software.

New products

Establish requirements, constraints, decisions, and acceptance criteria before or alongside implementation. Keep the ratified contract current through reviewed changes.

Existing systems

Code, tests, documents, runtime observations, and expert knowledge are evidence of the system. They do not automatically define intended behavior. Resolve conflicts and ratify one bounded area before a consequential change.

Loft connects specialist-and-agent loops; it does not replace professional roles.

Product, architecture, engineering, and assurance specialists can use agents suited to their work. Loft comes to Claude Code and Codex as one plugin, with its agent skills, its checks and Yard. Coding agents investigate, generate, test and change the code; the specification, the trace and every branch's Yard Log stay in your repository, apart from any agent transcript or model runtime.

Loft is built with Loft.

Loft's specification, its agent skills and plugins, its reconciliation, and Yard are specified and built with Loft itself, and its local edition is public and open source. An engagement takes the same loop to your product and your people.