The short answer: four stages, and a save decision you make on purpose

In AnsClaw Script Mode, you turn a repeatable Android task into a script by working through four stages on one master phone: Record, Compose, Test and Save. The save step is not a formality — before you name and keep a script, you check two separate things: what the test report says, and what actually happened on the phone screen. A saved draft that has not been verified is not a passed recording, and it is not a reliable script you can hand to a group of devices.

That distinction matters most to operators running a physical Android phone farm from a Windows host. The expensive mistake is not a failed recording; it is a script that looked fine in a report, got saved, and then went out to selected devices as if it were finished. This walkthrough follows the manual's Script Mode stages so you can make the save-or-re-record call deliberately the first time.

Before you start: bind one master phone and set a clear goal

Script Mode begins with a single, explicitly chosen device. A normal AI conversation in AnsClaw targets the open master-phone window or a device you have explicitly bound — it does not broadcast one conversation across every selected device. If no master or binding is in place, tools can reject the request rather than silently picking a phone for you. So the first practical step is to open or bind the one phone you intend to record on, and confirm you are looking at that device.

Then set the goal in plain operational terms. Not "manage the app" but something like: open this app, reach this screen, tap this identifiable non-destructive control, and confirm the resulting state. A goal you can describe as a sequence of observable screen states is far easier to record, test and judge than a vague intention. Write it down before you touch Record — you will use it again as your pass criteria.

Record: let the AI carry out the task on the master phone

With the master phone bound and the goal set, the AI assistant enters Script Mode and performs the task on that one device. The visible action is the point: the agent inspects the current screen, navigates or taps an identifiable non-destructive element where permitted, and then inspects the resulting state. You are watching a real task happen on a real phone, not a description of what could happen.

This is where you should be honest with yourself about what you are capturing. If the screen state after a tap is not what you expected, stop and correct the goal or the starting condition now. If a UI state or permission is uncertain, pause and confirm it yourself rather than letting the recording continue on an assumption. Recording a slightly wrong path just moves the problem into Compose and Test, where it is harder to see.

Compose: shape the captured steps into something repeatable

Compose is where the captured actions become a structured, repeatable task rather than a one-off trace. Think of it as editing the draft: the sequence should read as the same path you would take by hand, with the same starting screen and the same identifiable targets. If the recording picked up an incidental detour, this is the stage to tidy it.

Keep the scope modest on your first pass. A short, stable task with a clear end state is much easier to test and to trust than a long chain of steps with several branch points. Conditional AI recovery is not guaranteed, so do not design a draft that only works if the AI rescues an unexpected popup mid-run. If the task genuinely depends on handling an exception, plan for a human to review that case rather than assuming the script will cope.

Test: read the report and the phone, not just one of them

Test is the stage that earns the save. Run the draft and then look at two things side by side: the test report and the actual on-device result. The report tells you what the run recorded; the phone screen tells you whether the business task actually completed. These can disagree, and when they do, the phone wins.

This is the same principle mature device-testing operations apply at scale — a device showing activity is not evidence that a job succeeded, which is why serious test farms report per-device run and job outcomes rather than relying on a passive screen wall. You are applying that idea earlier, at the recording stage: the outcome you can observe is the outcome you trust.

If the report and the screen agree that the task completed as intended, you have something worth saving. If either is ambiguous, re-record or re-compose. Do not talk yourself into a pass because the report looks clean while the screen shows a stuck state.

Save: name it only after the result is verified

Save is the commitment step. Once the test report and the real phone result line up with the goal you wrote down at the start, name the script clearly — what app, what task, what starting condition — and save it. That name is what future you, or a colleague, will rely on when deciding which script to run on which devices.

Be precise about what a save does and does not mean. A saved script that was never verified is still an unverified draft; saving it does not retroactively make it a passed recording, and it does not make it a reliable script. The manual's limit is explicit: review both test results and the actual outcome, because an unverified script is not proven reliable. Treat Save as the moment you certify a result you have seen, not the moment you file a hopeful draft.

What happens after a verified save

A verified saved script is the reusable asset. It can then be run on several selected Android devices, with each device producing its own separate result. That is the practical payoff of doing the four stages properly: you are not repeating the same manual taps on every phone, and you are not guessing whether the run worked.

Replay is also where the economics of AI phone farm automation show up. Replaying a tested script does not require the AI to re-plan the task from scratch each time. One supplied single-device local test on AnsClaw 2.8.0.0 with a vivo Android 10 phone illustrates the shape of this: one standard AI run took 84.75 seconds and 396,136 tokens with a nonzero exit, followed by three successful normal replays of an unverified draft averaging 8.48 seconds with zero tokens each. Read that carefully — it is one short, stable browser task, the raw logs were not independently checked, the P1/P3 recording checks failed, and the replays did not invoke AI correction. The observed ratio is not a multi-device benchmark, not proof of other-app compatibility, and not proof of actual billed savings. It is a reason to test and verify before you scale, not a promise.

Handling exceptions across a group of phones

The common overreach in phone farm software marketing is the idea that you can ask the AI to scan every phone, find the one exception, and fix it. The nearest true workflow is more grounded. Each selected device produces its own result from the replay. You, the operator, review those per-device results, identify which phone needs attention, open or bind that affected phone, and then the AI may help inspect that screen within its authorized tool limits.

That human decision point is a feature, not a gap. It keeps a single unexpected popup from becoming an invisible failure across the farm, and it keeps the AI working on one explicitly selected phone at a time rather than acting blindly on many. Build your review habit around per-device outcomes from the start, and the exception cases become a short list to work through rather than a mystery.

A first-use checklist for your first script

Bind or open exactly one master Android phone and confirm you are looking at it. Write the goal as observable screen states, including the starting screen and the end state. Enter Script Mode and let the AI perform the task on that device, watching each step. Pause for your own confirmation if a UI state or permission is unclear. Compose the draft into a short, stable sequence with identifiable targets. Run Test, then compare the test report against the actual phone screen. Save and name the script only when both agree with your goal. Then replay on selected devices and review each device's separate result.

If you are evaluating AI phone farm automation for a physical Android fleet managed from Windows, this four-stage loop is the honest unit of work to judge. Ask whether the tool lets you see the action on one master phone, whether it gives you a test result you can compare against the real screen, and whether it keeps per-device outcomes visible after replay. Those three questions tell you more about day-to-day reliability than any feature list.