Start a local project with Codex or Claude Code on desktop

You already use chat to draft text. Now you want the result saved in files: a simple FAQ page, a reusable document or a small internal tool. A desktop agent adds a working loop around that request. It can inspect a selected project, propose changes, edit files and use available tools. Your job becomes setting the scope and reviewing the result. This guide uses an original fictional desk-repair business, so the first exercise needs no customer records, connected accounts or website credentials. The target is a local HTML page you can open yourself. It is a proposed learning exercise, not a report of a completed product test.

Before you start

• An approved account with access to the chosen coding features; check current entitlements and your organization’s rules.
• The current official desktop application for your supported system, a browser and a new empty practice folder. Use the linked installation guidance rather than a third-party download.
• The downloadable fictional project brief below, saved as project-brief.txt in that practice folder.

Tools: Codex, Claude Code. Use an assistant approved by your organization. This text-based first version does not require new integrations; account features and data policies still apply.

The workflow, step by step

1. Choose the desktop surface that matches the job

   OpenAI’s current quickstart places Codex in the ChatGPT desktop app: choose Codex from the ChatGPT dropdown for software development. ChatGPT also offers Chat and Work for other kinds of assistance. Anthropic documents Claude Code in the Claude desktop app’s Code tab. These are product-specific routes, so follow the current official instructions if your screen differs. For this exercise, choose the coding surface because you want an HTML file changed inside a folder. A desktop interface suits someone who prefers selecting projects, reading a conversation and inspecting changes visually. A terminal CLI becomes useful when you already work through commands; it is not a required upgrade.

2. Confirm access without mixing up the bill

   Sign in through the application’s own account flow and check the active account or workspace. For Codex, OpenAI distinguishes ChatGPT subscription access from API-key access; API usage has separate billing. Claude Code also has subscription and API/provider authentication routes. A paid chat account does not establish identical features or allowances across these products. Check your own access before buying anything. This page needs no model API, deployment account or secret key. Do not paste passwords, login tokens or API credentials into the prompt. If a work administrator controls the feature, use that approved route instead of connecting a personal account to company material.

3. Make the project folder concrete

   Create a folder named cedar-faq-practice in your normal documents location and save the brief inside it. A project folder is simply the directory containing the files for this task. Open or select that folder in the app and ask the agent to confirm its path and list its contents. Selecting the right directory matters more than giving the chat a memorable title. Keep unrelated documents elsewhere. Read any existing project instructions first: Codex documents AGENTS.md, while Claude Code documents CLAUDE.md. For a brand-new exercise the downloaded brief can carry the task rules; do not replace an existing project’s instruction files just to follow this tutorial.

4. Ask for a plan you can actually review

   Paste the starter prompt and request a short plan before edits. The agent should identify the input brief, the proposed index.html output and review-notes.txt. The page contains five answers based only on approved facts. Quote after inspection stays a quote, and undecided warranty terms stay undecided. Ask it to explain HTML as the page structure and CSS as its presentation if those terms are new. Reject an approach that adds a booking service, database or website framework: none is needed for this task. A useful plan is small enough that you understand each proposed file and how you will decide whether it is finished.

5. Use permissions to support the agreed scope

   Review the active permission settings before proceeding. A sentence in a prompt guides behavior; it does not create an operating-system boundary. Use a setup that lets you review edits and commands while learning. Anthropic’s current desktop documentation describes a Manual mode for this, and OpenAI documents approval and sandbox controls. Labels and defaults can vary. If the agent requests a command, ask what it will read, write and contact. A terminal is a tool for executing instructions on the computer, even when launched from a friendly app. The plain HTML exercise requires no package installation or broad machine access. Keep those requests outside the agreed task.

6. Inspect the file changes before judging the appearance

   After approving the plan, let the agent create the two output files. Open the change view or request a diff: a comparison showing what was added, removed or altered. Read the business answers first. Check that appointments remain Tuesday–Friday, that Harbor District is labelled fictional, and that no price or promise appeared from nowhere. Then inspect structure: one main page heading, five clearly separated questions and readable text. Ask the agent to explain any unfamiliar addition in ordinary language. A polished screenshot cannot show whether the source includes an unwanted form or external tracker. Review the actual files as well as what the page looks like.

7. Preview locally and record what was checked

   Open the saved index.html in your browser. This brief uses inline styling and no external dependencies, so opening the file directly is an appropriate preview. Narrow the browser window, increase text size and use the keyboard to inspect any interactive elements; this page should need none. Verify all five questions against the brief and read the fictional label. Ask for review-notes.txt to distinguish checks the agent performed from checks you performed and checks still pending. Automated tests, a successful file write and a visual preview answer different questions. If no browser check happened, record that honestly. There is no reason to publish the page to complete this exercise.

8. Make one revision and leave a clear handoff

   Request one bounded improvement, such as shortening the estimate answer while preserving its meaning. Review that diff again and reopen the file to confirm the change. Keep the original brief, final page and review notes together, with a backup before further experiments. If you later want a real business page, start a separate task with verified facts, an appropriate project and explicit publication scope. The transferable habit is straightforward: give the agent a concrete input, define the files it may change, inspect its work and check the outcome yourself. Once that loop feels comfortable, a larger project or the CLI guide becomes easier to approach.

Try it with a sample

Use this synthetic input to see the distinctions your output should preserve.

What a good result should preserve

A prompt you can adapt

Read project-brief.txt in the selected practice folder. First confirm the folder path, list existing files and summarize relevant project instructions. Explain a short plan and stop before editing.

After I approve that plan, create the local FAQ page exactly to the brief. Edit only index.html and review-notes.txt. Preserve all existing unrelated files and instruction files. Use plain HTML with inline CSS, no JavaScript, dependencies, forms, tracking or external assets. Do not connect accounts, deploy, publish, install software or access anything outside the practice folder.

Show the changed files and explain their diff. Check all five required answers against the brief, including unknown price and warranty. Report actual checks separately from suggested/manual checks. Do not claim browser preview, accessibility compliance or tests passed unless supported by what you actually did. If a command or permission request exceeds this scope, explain the need before proceeding. Provide a local file path I can open and a concise review checklist.

Check the result

• Folder: the selected directory contains only the intended practice project; unrelated files and instructions remain unchanged.
• Content: five answers match the fictional brief; no invented prices, warranties, contact details or live booking link.
• Changes: inspect both output files and explain unfamiliar additions before accepting them.
• Preview: open index.html yourself and check narrow-window readability; record the actual result.
• Scope: no publication, account connection, dependency installation, external asset or secret appears.

What to keep in mind

This is official-source research plus original synthetic teaching material. We did not install these products, sign into reader accounts or run this FAQ through either agent. Interface, platform support, permissions and access can change; the linked official pages establish the reviewed guidance, not a promise about your account. A local preview is neither production readiness nor a complete accessibility or security audit.

Sources & review notes

These sources support the tool context and review principles. The workflow, prompt, and synthetic example are original editorial material.

• OpenAI: Quickstart

  Current desktop instructions select Codex for software development; ChatGPT also has Chat and Work modes. Follow current downloads for system requirements.

• OpenAI: Authentication

  Distinguishes ChatGPT subscription sign-in from usage-based API authentication and its separate account policies.

• OpenAI: Custom instructions with AGENTS.md

  Documents project instruction discovery; existing instructions should be inspected before adding task guidance.

• Anthropic: How Claude remembers your project

  Documents CLAUDE.md project instructions and their scope; this is guidance, not a filesystem permission system.

• OpenAI: Agent approvals and security

  Documents sandbox and approval controls; actual permissions depend on the active configuration and workspace policy.

• Anthropic: Authentication

  Documents subscription, Console and provider authentication. Check the active route before assuming which account pays.

• Anthropic: Desktop application

  Describes the Code tab, project sessions, permission modes and panes for reviewing changes and terminal activity. Interface details depend on version.