Tutorial

Run product development with your coding agent

By the end of this tutorial you will have taken one feature from a sentence to shipped work, with Claude Code or Codex doing the work and Agiflow holding the state. Five prompts, one board, nothing kept in a chat thread.

01 : What you will have at the end

A board where every task carries acceptance criteria, your agent moved each one through the status flow, and the review verdict is recorded against the criteria it was measured on. You can close your terminal, come back tomorrow, and the agent picks up exactly where it left off.

Agiflow project board with tasks distributed across the Planning, Todo, In Progress, Testing, Review, and Done columns. — The finished board. Every column carries real work, and the agent moved it there.

Status flowPlanningTodoIn ProgressTestingReviewDone

Two exception paths exist: Blocked when a human has to step in, and Cancelled when work is dropped.

02 : Before you start

An Agiflow project

Free plan is enough. You need one project you are willing to put real work into.

Claude Code, Codex, or ChatGPT

Any one of them. The board is the same; only the setup command differs.

The AgiFlow AI Plugin

The bundle that gives your agent the five workflows below. Installed in the next section.

03 : Install the plugin

The AgiFlow AI Plugin is a single bundle that wires the MCP connection and adds the workflows this tutorial uses. Pick your runtime. Everything after this section is identical across all three.

Install and connect

git clone https://github.com/AgiFlow/ai-plugin.git agiflow-ai-plugin
claude --plugin-dir ./agiflow-ai-plugin

Config lives in: .mcp.json
Check it worked: /mcp
Run a workflow: /agiflow:run-task

Install and connect

codex plugin marketplace add AgiFlow/ai-plugin
codex plugin add agiflow-ai-plugin@agiflow
codex mcp login agiflow

Config lives in: ~/.codex/config.toml
Check it worked: codex mcp list
Run a workflow: Ask for the workflow by name, for example "run task DXX-2".

Install and connect

Add the Agiflow app from the ChatGPT App Store, then authorize one project.

Check it worked: Ask ChatGPT to list your Agiflow projects.
Run a workflow: Ask for the workflow by name, for example "plan this project in Agiflow".

💡 Use the plugin endpoint, not the widget one:

The bundle points at the coding-agent MCP endpoint. The versioned widget endpoint serves the ChatGPT app surface and will not expose these workflows to a terminal agent. If you configured Agiflow by hand before, check the URL in your config before continuing. Full connection detail is in supported AI tools.

Claude Code terminal session showing the Agiflow MCP server connected and its available workflow skills listed. — Claude Code with the plugin loaded. The Agiflow server is connected and the workflows are available.

04 : The loop

Five workflows, run in order. Each one owns a specific transition, and none of them skip ahead: a task cannot be executed until grooming has promoted it, and it cannot be closed until review has judged it.

project-plan

Turns a goal into small, testable tasks with acceptance criteria.

creates tasks in Planning

refine-task

Makes one vague task unambiguous without widening its scope.

stays in Planning

backlog-grooming

Checks readiness, groups tasks into work units, and orders what is next.

Planning to Todo

run-task

Executes one task against its acceptance criteria, with tests.

Todo to In Progress to Testing to Review

review-work

Gives every acceptance criterion a verdict and files follow-up tasks.

Review to Done, or back to In Progress

project-plan

Step 1 : Turn a goal into tasks

Send this

Plan the work for adding passwordless email sign-in to our web app. Keep tasks small enough to finish in one session.

You should see

Your agent reads the project, checks for overlapping tasks, asks the few questions it genuinely needs, then proposes a task list and waits for your approval before writing anything.

Agiflow board Planning column containing newly created tasks, each showing a priority badge and an acceptance criteria count. — Tasks land in Planning, not Todo. Nothing is executable yet, and that is deliberate.

Each task carries at least two acceptance criteria that can be checked objectively.
Work units are not created here. Grouping happens during grooming, once readiness is known.
If the agent proposes a task that bundles two unrelated outcomes, ask it to split them before you approve.

refine-task

Step 2 : Make one task unambiguous

Send this

Refine the passwordless sign-in token expiry task. I want the acceptance criteria to be measurable.

You should see

Your agent reads the task and its comments, names the ambiguities, asks about the ones it cannot resolve, and rewrites the criteria. Vague phrasing such as "handles errors correctly" is replaced with observable behaviour.

Agiflow task detail view showing a refined description, explicit in-scope and out-of-scope boundaries, and a list of measurable acceptance criteria. — The same task after refinement. Scope boundaries are explicit and every criterion is checkable.

Refinement never adds new requirements. If the agent starts inventing scope, that is a bug in the prompt, not a feature.
The task stays in Planning. Only grooming promotes it.

backlog-grooming

Step 3 : Decide what is actually ready

Send this

Groom the backlog. Promote what is ready and group the sign-in work into a work unit.

You should see

Your agent classifies every Planning task as ready, needs refinement, blocked, or duplicate. It proposes a priority order with the tradeoffs, suggests work units, and asks for approval before it writes.

Agiflow board showing a work unit grouping several related tasks, with ready tasks promoted into the Todo column. — Ready work is promoted to Todo and grouped into a work unit. Anything that failed the readiness check stayed behind.

A task missing testable acceptance criteria is never promoted. Send it back through refine-task.
A work unit holds three to eight tasks that deliver one shared capability. One or two related tasks stay standalone.
Downstream tasks whose upstream work is incomplete stay in Planning, gated on the dependency.

run-task

Step 4 : Execute

Send this

Run task DXX-2.

You should see

Your agent moves the task to In Progress, works through the acceptance criteria one at a time, records which files changed and what the tests returned, then moves it to Testing and finally to Review with a draft commit message and PR body attached.

Agiflow task detail view in In Progress status with acceptance criteria checked off and a progress comment listing changed files and test results. — Progress is written to the task as it happens, so a new session can resume without re-reading the conversation.

A task still in Planning is refused. That guard is the reason grooming exists.
When tests fail twice, the task goes to Blocked with a comment explaining what a human needs to do. It does not loop forever.
The board holds the state, so you can stop the session at any point and pick it up on another machine.

review-work

Step 5 : Judge the work

Send this

Review work unit DXX-WU-1.

You should see

Your agent gives every acceptance criterion a pass, fail, or partial verdict with the evidence it used, flags issues by severity, and either approves the work or sends it back with specific instructions.

Agiflow task comment containing a review summary with a verdict, per-criterion results, and a list of issues grouped by severity. — The verdict is recorded against the criteria it was measured on, not as a general impression.

Review does not fix anything. Gaps become follow-up tasks so the record of what shipped stays honest.
A pass with known gaps is worse than a clear needs-rework. Expect the agent to say so.

05 : When it goes wrong

⚠️ The agent cannot see your project:

Almost always the MCP URL. Coding agents need the plugin endpoint, not the versioned widget endpoint. Re-run the check command for your runtime from section 03, then re-authenticate.

⚠️ The agent invents status names:

Your board may use custom columns. The workflows read the real status names before writing, so if you see invented ones, the agent skipped that read. Ask it to list the project statuses and try again.

⚠️ Authorization loops or never returns:

Close stale Agiflow sign-in tabs and start again. More detail in troubleshooting connections.

06 : Where to go next

The same bundle carries workflows for the parts of the week that are not writing code.

getting-startedReads the workspace and recommends what to do next.

daily-standupRead-only progress check across the board.

triageFinds stalled, blocked, or overloaded work and unblocks it.

orchestrateCoordinates several agents against one board.

run-workExecutes a whole work unit instead of a single task.

Share this with your team

The same walkthrough as a deck, with a PDF you can attach to an email.

Open the deck

The bundle is open source at github.com/AgiFlow/ai-plugin. Clone it as agiflow-ai-plugin and read the skills directly if you want to see exactly what your agent is being asked to do.