Claude Code project management

Claude Code project management starts from an ordinary observation: Claude Code is a terminal agent, and a terminal session knows nothing about your backlog. You open it in a repository, describe a task in your own words, and it works. That is excellent for a single focused change and awkward for a project with forty open items, three of them blocked and one waiting on a decision. You end up keeping the plan in a notes file or your head, retyping context, and trusting the agent's own statement that it finished. A project workflow puts a record in front of the session. The record says what the task is, what files it may touch, what done looks like and who reviews it, and the session becomes one step in a loop that ends with a human decision. This page walks that loop and describes what Liquid Brain adds to it.

How Claude Code behaves on its own

Claude Code runs in your terminal against your working directory. It reads files, proposes edits, runs commands and asks permission for actions according to how you configured it. It also offers a plan mode in which it reasons about a task without making changes, which is a good way to see its intended approach before anything is written. A project instruction file is read at the start of a session. What it does not provide is a project view. It has no built in notion of which tasks exist, which are ready, or which finished yesterday. Each session is a fresh conversation unless you resume one. That is by design: the tool is a capable worker, not a planner. Anything that organizes work across sessions has to come from outside it, which is where a card, a board or a harness enters the picture.

A workflow around the session

A dependable loop has four steps. First, write the task as a record with a clear outcome, scope and completion criteria. Second, start the session with that record as its context, ideally in a read-only or planning stance first, so you can check the approach. Third, let the agent do the work inside declared boundaries. Fourth, inspect the result against the criteria and accept or reject it yourself. The agent finishing is an event, not a verdict. A clean exit code tells you the process ended, and a diff tells you what changed, but neither tells you the change is correct. The discipline is to keep the acceptance decision with a person and to record what ran, so a later session or reviewer can see the attempt rather than reconstruct it from a scrollback buffer.

Running Claude from a Liquid Brain card

In Liquid Brain you can open Claude from the card itself, with its working context close at hand [C1b]. A card offers a start action with a project folder and a choice of agent, and a read-only style assessment option exists for reading context and proposing next steps before any implementation [C1b]. Explicit scheduled tasks can run review or work, with readiness and authority checks before execution [C1c]. Proposed scope stays separate from approved execution authority [N09]. Configured command guards and scope controls check work against declared boundaries [C5]. Afterwards, recorded attempts and outcomes stay with the card [C4], and accepted harness work lands in a review state rather than done by default [C4]. All of this requires a configured, compatible Claude CLI on your computer [C1b]. Liquid Brain organizes and constrains the work. Claude still performs it.

QUESTIONS

Does Liquid Brain replace Claude Code?

No. Your configured Claude CLI still does the work, and Liquid Brain organizes the cards, supplies context and records outcomes around it. You need Claude Code installed, configured and working on its own before a card can open a session from it [C1b]. Nothing here changes how Claude Code itself behaves inside the session.

Can Claude Code run on a schedule?

On its own it is an interactive terminal tool. Liquid Brain lets you schedule explicit review or work tasks with readiness and authority checks [C1c], but the Desk must be running, and a terminal exit is not proof the work was completed.

How do I review what Claude Code changed?

Compare the change against the card's completion criteria, not only the diff, because a diff is not proof of correctness. Check that the touched files stayed inside the declared scope, look for the evidence that the relevant check ran, and then decide. Liquid Brain keeps recorded attempts and outcomes with the work so you can inspect them before accepting [C4].

RELATED GUIDES

• Cards vs scheduled tasks: what starts an agent?
• How to preview work with the Liquid Brain harness CLI
• How to create a project card in Liquid Brain Desk

RELATED HARNESS PAGES

• Codex CLI project management Codex CLI project management as a workflow: launch Codex from project cards, schedule explicit tasks and inspect outcomes before accepting.
• human review of agent output Human review of agent output: what to inspect before accepting AI changes, and how recorded outcomes make the review honest.
• scheduled agent tasks Scheduled agent tasks done carefully: explicit prompts, readiness and authority checks, and why a started run is not a finished one.