AI writes working PowerShell in seconds — including working PowerShell that runs as admin, swallows errors, and deletes the wrong thing at scale. The skill this playbook teaches is not prompting for scripts; it is the fixed five-point safety review that every generated script passes before it touches a real machine.
Tier: Core · Workflow area: Scripting & automation · ~35 minutes
This playbook establishes the generation-plus-review protocol for admin scripting with Claude, Copilot or Gemini: a prompt pattern that requests safety features up front, then the mandatory review — -WhatIf support, least privilege, error paths, secrets handling, PSScriptAnalyzer — applied every time, with no seniority exemption.
Get-ADUser jsmith -Properties LockedOut needs no protocol. The line is state change and scale, not script length.Prompt for safety features, not just function. The generation prompt bakes in the requirements so the review starts from a defensible draft:
Write a PowerShell script: disable AD user accounts inactive for 90+
days, from OU "OU=Staff,DC=corp,DC=local", excluding accounts in the
group "ServiceAccounts".
Requirements — non-negotiable:
- [CmdletBinding(SupportsShouldProcess)] and honor -WhatIf/-Confirm
on every state change
- Parameters with validation, no hardcoded values
- Try/catch on every AD call; on error: log and continue, never
silently skip; summary of failures at the end
- No plaintext credentials anywhere; assume the runtime identity
- Before changing anything, export the affected accounts (name,
DN, lastLogonTimestamp) to a timestamped CSV — the rollback record
- Log every action taken to a transcript file
Then list: what could go wrong running this, and what permissions
it minimally needs.
That final question is a review head start — the model is decent at criticizing its own script when asked explicitly.
Review point 1 — dry run honestly supported. Confirm SupportsShouldProcess is declared and every state-changing call actually respects it (if ($PSCmdlet.ShouldProcess(...)) or the cmdlet's native -WhatIf passthrough). A script that declares it but bypasses it in an inner function is the classic generated bug.
Review point 2 — least privilege. Ask: what is the minimum identity this needs? Generated scripts assume Domain Admin by default silence. Run it as a delegated service account scoped to the OU or Graph scope in question; if the script fails under least privilege, fix the scoping, not the privilege.
Review point 3 — error paths. Read every catch block. The generated failure modes to hunt: empty catch (error vanishes), catch { continue } without logging, and a pipeline where one failed item kills the whole run. Force a failure to test it — point the script at one nonexistent object and watch what it does.
Review point 4 — secrets. Grep for password/key/token literals, connection strings, and ConvertTo-SecureString ... -AsPlainText. Anything found moves to the platform's secret store (SecretManagement module, Azure Key Vault, RMM credential store) or dies.
Review point 5 — run PSScriptAnalyzer. Mechanical and non-negotiable:
Install-Module PSScriptAnalyzer -Scope CurrentUser
Invoke-ScriptAnalyzer -Path .\Disable-StaleAccounts.ps1 -Severity Warning,Error
Typical catch on generated code:
RuleName Severity Line Message
PSAvoidUsingPlainTextForPassword Warning 34 Parameter '$Password' should use SecureString
PSUseShouldProcessForStateChanging Warning 61 Function calls Set-ADUser without ShouldProcess
Errors block; warnings are resolved or explicitly waived in a comment with a reason.
Execute in stages, keep the artifacts. -WhatIf first and read the full output — the WhatIf list is your blast-radius preview. Then limited live run (-First 5 or a pilot OU), verify in the console, then full run. File the WhatIf output, the pre-change CSV and the transcript with the change record.
PULL_REQUEST_TEMPLATE so scripts merged to the team library carry evidence of all five points.--dry-run flag you implement, ShellCheck instead of PSScriptAnalyzer, set -euo pipefail as the error-path baseline, and secrets via your vault, never in the script.Never paste real credentials, tenant IDs or production hostnames into a public assistant while iterating on a script — use placeholders, and populate real values only in your editor from the secret store.
Generation is the easy 10%; the fixed five-point review — dry run, least privilege, error paths, secrets, analyzer — is what makes AI scripting an asset, and it applies to every script, every time, regardless of who or what wrote it.