SKILL PLAN · DETERMINISTIC

A small plan for a real job

Start with the minimum useful stack. Add nothing unless it has a clear responsibility in the run.

Task
Web app testing
Agent
Codex
Priority
Balanced
Start over

Rules + reviewed metadata only. No LLM. Same inputs produce the same plan.

03 SKILLS · Install order

Your Skill Plan

Install and verify one role at a time. The plan never forces three Skills when fewer are enough.

01Core execution

Native · Codex

playwright-workflows

Exercises the real user flow in a browser.
Open profile and install guide
Review permissions, limitations, and evidence
Agent fit
Native · Codex
Main limitation
Detailed package files are not fully catalogued.
First risk
Review before install · Medium
Source
Editorial source
02Guardrail

Native · Codex

web-design-guidelines

Defines scope and prevents a broad, unfocused test run.
Open profile and install guide
Review permissions, limitations, and evidence
Agent fit
Native · Codex
Main limitation
It depends on network access and a live guideline document whose contents can change over time.
First risk
Rule drift or context-free findings · Medium
Source
Pinned source · b8caa260
03Evidence

Native · Codex

minimal-run-and-audit

Retains visual and runtime evidence for review.
Open profile and install guide
Review permissions, limitations, and evidence
Agent fit
Native · Codex
Main limitation
It does not choose the reproduction target, train a model, conduct broad paper analysis, or hide scientific-meaning changes.
First risk
Scientific changes hidden as setup fixes · High
Source
Pinned source · 3ab50525

Combined permission budget

This is the union of declared permissions across the plan. Review each package before installation.

Browser controlRun commandsWrite screenshotsRead frontend filesReview package instructions before install

Finish with evidence

Prove the critical user flow works in a real browser and leave repeatable evidence.

Use test data and a local or explicitly approved environment.

  1. 3-step acceptance check
  2. Run one critical flow from a clean browser state.
  3. Assert the expected page state, not only a successful click.
  4. Save the failure signal needed to reproduce any defect.
  • Stop conditions
  • The test requires production credentials or personal data.
  • The script performs destructive actions without isolation.
  • A flaky result is reported as a pass.