Wizard

Wizard

Turn one settled human-only procedure into one guided local script. The script may direct browser or physical actions and collect values at runtime. The authoring agent never receives real secrets or participates in the manual, credentialed, destructive, or external-mutation stages.

The invocation authorizes one local script inside the requested scope. When the same request clearly asks to begin, run, or be walked through the procedure, it also authorizes launching the verified script for the human in an isolated visible terminal. It does not authorize installing a runtime or dependency, committing, publishing, or changing unrelated documentation.

1. Qualify

Inspect the repository, available tools, and requested result. Separate work the agent can perform with current authority from steps that require the human's identity, credentials, judgment, physical access, or dashboard session.

Use Wizard only when at least two dependent settled stages require the current human and scripting materially reduces secret exposure, copy-paste mistakes, target-selection mistakes, or ordering mistakes. Otherwise perform authorized work in its current owner or return the one manual action directly. Return an unresolved decision or missing fact to its owner instead of encoding a guess.

2. Map

Name the stages in dependency order. For each one, establish its authoritative instruction source, human action, and destination or effect. When a stage captures a value, also establish its secret status. When it mutates state, establish the exact local or external target. Inspect examples, configuration, consumers, and variable names without reading existing secret values. Never ask the user to paste a secret into chat.

Check current official instructions when a third-party command or interface may have changed. Ask only when a consequential instruction, target, destination, or authority remains unknown. Do not treat every referenced environment or CI name as Wizard-owned.

3. Author

Write one script in the requested location. Otherwise use the repository's ignored scratch convention, or .tmp/wizard-<slug>.<ext> when none is defined. Choose an existing repository runtime or the operator's native shell. Add no runtime or dependency for the wizard. Use a durable scripts directory only when the user requested a repeatable repository workflow.

Keep each stage focused. Open or print the authoritative source or instruction before asking the human for a value. Identify the destination without displaying its existing contents. Capture secrets without echoing or embedding them. Before writing a secret-bearing local file, resolve its exact path and establish that the repository intentionally excludes it from source control.

Use native secure input or standard input when a tool supports it. Never place a secret in command arguments, logs, transcripts, or temporary files.

Before a destructive action, show its exact target and confirm. Before a durable external effect, show its exact target and active identity, confirm authority, establish an observable postcondition, and read it back afterward. When a secret value cannot be returned, verify its exact scope, name, and fresh metadata and state that value equality remains unproved. Leave the mutation manual and unverified when no useful postcondition is observable.

Stop dependent stages after a required stage fails or is cancelled. If an effect may have partially succeeded or its result is indeterminate, inspect current state before choosing recovery. Make an expected resume or retry converge without duplicate effects. Leave harmless stages free of those controls, and never report completion while required work remains unfinished.

4. Check

Run the nearest syntax or lint check for the chosen runtime. Statically trace every declared input to its destination and every effect to its target, applicable confirmation, read-back, and applicable recovery path.

When safe, execute non-credentialed local helpers against an isolated dummy target under the repository's ignored scratch convention. Exercise the creation, update, cancellation, and read-back paths the script actually supports. Do not open a browser, request credentials, alter real configuration, or perform destructive or durable external mutation during these checks.

State what the checks prove and what the operator's first run must still verify. If the invoking request already asks to begin, launch the checked script in a separate visible terminal whose input and output are not attached to the agent. Do not ask for another confirmation merely to open it. Pass no secret in its arguments or environment, and do not read, poll, or capture the terminal. If that separation is unavailable, return the run command instead.

5. Return

Return the script path, one run command, the effects it may perform, the checks run, whether it was launched, and any unresolved manual detail. Stop before installation, commit, publication, unrelated cleanup, or another workflow.

Complete when one secret-free, checked script covers every admitted human-only stage and the run command fits the target environment. When launch was requested, start the isolated terminal or return the exact separation gap. The human-owned procedure and its external outcome remain unverified.