Before running a script written by Swarmbot AI, ask these three questions
A draft from Swarmbot's script assistant is a starting point, not a run approval. Check its device and page scope, possible changes, and required inputs or image assets before testing one small step.
Review a draft from Swarmbot's script assistant before you run it. This matters when you are new to AI-assisted scripting or plan to hand the script to a teammate. Open the current project and its assistant in Swarmbot's script development view. Keep your original task description beside the draft, and ask AI to explain in plain language what it proposes to do. Code that looks complete is not a reason to send it to several devices.
1. Where will it start, and which devices could it touch?
Compare the draft's assumptions with the device and app page in front of you. Is the app open? Does it require a signed-in state? Is the target actually on this page? Ask AI to identify the current Swarmbot script functions it used, the intended device scope, and the page cue for each step. An example device number, private file path, or old screenshot should not become a general rule.
Try this request in the script assistant:
Do not run anything yet. In five sentences, explain this script's starting page, device scope, the page cue for each action, and where it stops if a cue is missing. List uncertain prerequisites separately. Do not guess a fixed coordinate.
If AI cannot explain the starting state, first provide a cropped, redacted image of the current page or prepare one test device. Adding more code will not resolve an unknown starting point.
2. Will it only read, or will it change something?
Opening a page or reading a status is different from sending, deleting, submitting, or installing. Ask AI to label each draft step as read-only or state-changing. For changes, identify the human approval point, the stop condition, and what a second run would do. An AI message saying “done” is not proof that the phone reached the expected state.
For example, add this boundary:
This draft should stop after checking the search results page. Remove any step that could submit, pay, delete, or send a message. If such an action is genuinely needed, list it as a separate phase requiring my review.
That is easier to check than “make it safe.” GitHub's guide to reviewing AI-generated code likewise recommends checking intent, invented APIs, and unexpected dependencies. For a Swarmbot script, use its own current runtime documentation and project files as the source of truth.
3. Which inputs and assets does it need?
A Swarmbot script project has a manifest and entry code, and may include image assets. Ask AI to list every run parameter and image file referenced by the draft, then locate each one in the project. Changing search terms belong in the run form rather than in hard-coded text. An image-finding crop must match the current page; a familiar file name alone proves nothing. The previous guides cover run parameters and preparing a target image.
Once all three answers make sense, save the draft and try only one small, reversible step on an authorized test device. Compare the screen, script output, and expected page. If they diverge, give AI the first point of difference and ask it to revise just that part. This keeps you in control of what you authorize the script to do.