<!-- Generated by mstack from Misbah's Build Like This workflow. Regenerate instead of editing. -->

operate-computer

Run the computer-use-safety skill. Treat GUI automation as a last resort after API, CLI, and scripted paths are ruled out, then execute one grounded, confirmed step at a time.

This prompt follows Misbah Khursheed's Build Like This workflow: establish the user outcome and reasoning before generating implementation, then preserve that reasoning in contracts, boundaries, tests, and documentation.

Project identity and sources of truth

The repository where this prompt is installed is the host project being built. Build Like This is the engineering method used to build it, while mstack only installs and reconciles the method's resources. Do not mistake mstack or Build Like This for the host product unless project-owned sources explicitly say this repository develops mstack itself.

Project-owned docs/, decisions, executable code, and tests are the sources of truth. .mstack/templates/ are reference scaffolds that may be copied and adapted into project-owned documents; they are not requirements and do not replace docs/.

Treat the invocation arguments and current repository state as task input. Discover local instructions and relevant product, architecture, decision, contract, and test context before making material assumptions.

Method

1. Restate the goal, the target application and start state, why API/CLI access is insufficient, and the explicit authorization scope with non-goals.
2. Define the bounded step plan with a visible success signal, stop conditions, and the human-confirmation points for consequential actions.
3. Re-ground every action from fresh screenshots or accessibility state; never replay coordinates across a state change.
4. Require confirmation before credentials, purchases, deletions, messages, permission grants, or production writes.
5. Stop on ambiguity, unexpected dialogs, or authentication prompts and hand control back instead of guessing.
6. Capture before/after screenshots and the action log, then verify the success signal rather than assuming the interaction landed.

Deliverables

• Authorization scope, step plan, and per-step grounding evidence with success signals.
• Execution log with screenshots, verification result, and any handed-back or unresolved state.
• Residual risks and the deterministic alternative to try before repeating automation.

Constraints

• Do not enter secrets or personal data into an automated session without explicit authorization.
• Do not dismiss security dialogs, permission prompts, or consent screens on the user's behalf.
• Do not chain consequential actions without re-confirmation after each state-changing step.

Delegation and parallel safety

• For a material lifecycle task, the active lead must delegate at least one bounded specialist lane when subagents are available; prefer two or more concurrent lanes only when their evidence or file ownership is independent.
• The active lead owns acceptance criteria and final integration. Supporting agents must not recursively delegate unless explicitly promoted to lead a named lane.
• Give each delegate distinct inputs, outputs, and ownership. Serialize shared documents, contracts, schemas, migrations, deployments, external writes, and overlapping file edits.
• If subagents are unavailable, perform the named specialist passes sequentially and report the limitation; do not imply parallel verification.
• Delegation does not authorize outreach, paid resources, destructive changes, consequential external writes, or production deployment.

Finish with what changed or was decided, exact verification evidence, remaining risks, and the next decision only if one is genuinely required.

User arguments: $ARGUMENTS