ChatGPT 5.6 Complete Guide: Prompts, Work, Codex, Sites, and a Personal Control Tower

Direct Answer

The best way to use the new ChatGPT app is not to connect every account and ask it to run your life. Start with one read-only project, give it a small source registry, teach it one output format, and keep every consequential action in draft or preview. Once that works, add a scheduled review, one Codex build, or one private dashboard.

Matt Wolfe's walkthrough is useful because it shows the whole range: Chat, Work, Codex, Sites, connected apps, a signed-in browser, scheduled tasks, personal context, software builds, and a private business dashboard. The durable lesson is narrower: context makes the agent useful, but permissions determine whether that usefulness is safe.

Video credit: Matt Wolfe, creator of FutureTools. Follow Matt on X. Disclosure: Matt states that this video was sponsored by OpenAI. This article was not sponsored by OpenAI and independently checks product claims against OpenAI's documentation.

Source Note

Matt says he used 6.8 billion Codex tokens in roughly two weeks, including 1.3 billion on July 11 and 2 billion on July 13. That is a creator-reported usage figure shown in a sponsored video. It is not an audited bill, an efficiency score, or evidence that every generated token produced accepted work. His prepared examples also took from a few minutes to around an hour; this was not a sequence of instant one-shot builds.

The official product spine is clear. OpenAI's July 2026 release brings Chat, Work, and Codex into one desktop app for macOS and Windows, with access and limits depending on plan and workspace. Work can use connected apps, files, browser access, Computer Use, Sites, and scheduled work. GPT-5.6 provides Sol, Terra, and Luna across supported ChatGPT, Work, Codex, and API surfaces.

Product availability moves quickly. Apps, actions, browser features, model selection, and workspace controls vary by plan, region, and administrator. Check the live interface and official help page before designing a workflow around a feature.

Link Map

The New App Is Five Work Surfaces, Not One Chatbox

The built-in browser, cloud browser, and Codex Chrome extension are different trust surfaces. The cloud browser is appropriate for public pages and cannot sign in. The built-in desktop browser can hold its own signed-in state on eligible plans. The Chrome extension can use an existing Chrome profile when that is genuinely necessary. Choose deliberately; do not treat them as interchangeable.

The Useful Operating Pattern

1. Define the outcome. A weekly plan, one-page brief, draft reply, tested feature, or private dashboard.
2. Register only required sources. Add Calendar before Gmail if calendar evidence is enough. Add one business folder before an entire Drive.
3. Default to read-only. Retrieval and analysis are safer than write access.
4. Require provenance. Ask which source and timestamp support each important claim.
5. Generate a reviewable artifact. A draft, plan, report, diff, preview, or proposed schedule.
6. Approve at the point of consequence. Sending, deleting, moving files, publishing, purchasing, and changing production remain human decisions.
7. Measure accepted work. Track time saved, corrections, rejected output, and risk, not token volume alone.

Prompt 1: Build a Read-Only Personal Assistant

Matt's version can consult Gmail, Calendar, Slack, Granola, Drive, second-brain files, journals, prior chats, social accounts, YouTube, FutureTools, and his private Control Tower. That breadth makes the demo compelling, but it is too much for a first installation. Begin with a source registry and an explicit no-action contract.

Create a project-wide read-only assistant for [work area].

Approved sources for phase 1:
- [calendar or task source]
- [one approved folder, app, or project]

Operating rules:
1. Retrieve only the minimum context needed for the current request.
2. Treat every source as read-only.
3. Never send, post, delete, move, rename, publish, purchase, or change data.
4. For important facts, show the source, date, and a short evidence note.
5. Separate sourced fact, inference, and recommendation.
6. If sources conflict or access is missing, say so clearly.
7. Do not add new sources or permissions automatically. Propose them for review.
8. Draft outputs for me to approve; do not execute external actions.
9. Do not expose private source content beyond what the requested output needs.

First, show me the source registry, data risks, and a five-task acceptance test.
Do not retrieve data until I approve that plan.

Run five low-risk tests before adding another source: today's schedule, one project summary, one source-backed question, one conflict check, and one deliberate request for a prohibited action. A useful assistant should refuse the prohibited action and offer a draft or preview instead.

Prompt 2: Teach It Your Voice Without Teaching It Other People's Words

Matt asks ChatGPT to analyze messages he authored in sent email and work chat over the previous 12 months, excluding forwarded and copied text. He creates separate profiles for email and work chat, then uses them for drafts only. The separation matters: an email to a partner should not inherit the compressed style of a Slack message.

Build two writing profiles from messages I authored:
- email voice
- internal work-chat voice

Scope:
- use only messages written by me from [date range]
- exclude forwarded text, quoted replies, copied material, signatures,
  legal notices, confidential attachments, and other people's words
- do not infer sensitive traits or private facts

For each profile, identify:
- typical length and structure
- greeting and closing patterns
- directness, warmth, and formality
- sentence and paragraph rhythm
- phrases I genuinely use
- habits to avoid

Return a short editable style guide and three anonymized draft examples.
Use the profile only for drafts. Never send a message without my review.

Review the profile quarterly. Delete accidental imitations, outdated habits, and phrases that sound polished but not like you. A voice profile should be a constraint, not a costume.

Prompt 3: Run a Weekly Operating Review

The shortest useful prompt in the video is: What is on my calendar, Slack, and in my email that I need to do this week? Are there things you can take off my plate for me? The safer version preserves the convenience while preventing hidden actions.

Prepare my weekly operating review from [approved sources].

Return:
1. Fixed commitments with dates, owners, and source links.
2. Tasks that must finish before those commitments.
3. Decisions waiting on me.
4. Conflicts, missing information, and items that may be stale.
5. Work already completed, with evidence.
6. Three tasks you could help prepare as drafts or plans.

Rules:
- do not send, reschedule, close, or mark anything complete
- do not invent deadlines from message urgency
- label every item as confirmed, inferred, or needs review
- ask before creating reminders or scheduled tasks
- keep the main review to one page

One page is a feature. Matt's panel-preparation example produced 32 pages; that can be valuable as a source pack, but it is not the artifact you should carry into a meeting. Generate a concise brief first and keep the extended evidence as an appendix.

Prompt 4: Reuse Prior Context for a Meeting, Book, or Presentation

Matt points the assistant to earlier session IDs to create a one-page book pitch. He also combines a presentation outline with an event brief to generate a slide deck, then says he and a team member substantially improved the result. The model accelerated assembly; human judgment still owned the story and final design.

I am preparing for [meeting, panel, pitch, book, or webinar].

Approved context:
- [session or project reference]
- [event brief or outline]
- [specific source files]

Create:
- a one-page brief with objective, audience, three key messages,
  likely questions, and five relevant stories or examples
- an evidence appendix showing where each example came from
- a list of gaps I should answer myself

Do not expose confidential details, invent personal anecdotes, or quote private
messages. Keep the one-page brief concise even if the appendix is long.

For a slide deck, add brand constraints, audience knowledge, maximum slide count, source requirements, and a final checklist for factual review, image rights, legibility, and speaking flow.

Prompt 5: Research and Draft Outreach Without Turning It Into Spam

Matt demonstrates 20 San Diego small-business leads and 20 Gmail drafts for a fictional consulting offer. He does not send them and says he does not actually provide that service. The useful capability is research plus individualized drafting. The dangerous version is unreviewed volume.

Research five potential organizations for [specific service and region].

Qualification rules:
- there must be a current, public, business-relevant reason for fit
- use public professional contact routes only
- exclude prior opt-outs, existing customers, personal addresses, and weak matches
- record the source and date for every personalization claim

For each qualified organization, return:
- why it may fit
- evidence and source URL
- likely role to contact, without guessing a private address
- one useful observation or small no-cost idea
- a short draft email in my approved voice

Do not send, schedule, create CRM records, or create mailbox drafts yet.
Stop after five and ask me to review relevance and compliance.

After approval, the system may create drafts, but sending remains manual. Keep a suppression list, identify the sender clearly, honor opt-outs, and follow the rules that apply to the recipient's region. For the fuller workflow, see the AI-assisted cold-email framework.

For File Operations, Preview Before Moving Anything

The video shows Work organizing 65 downloads and renaming 21 thumbnails from their visual content. That is convenient, but file moves and renames can break project links, duplicate files, or make recovery tedious. Ask for an inventory and change manifest first.

Audit [folder] and propose an organization plan.

Phase 1 - read only:
- inventory file name, type, size, modified date, duplicate signals,
  and likely category
- flag ambiguous, sensitive, executable, hidden, and project-linked files
- propose destination paths and clearer names
- produce a change manifest with old path, proposed path, and reason

Do not move, rename, delete, overwrite, unzip, or run anything.
Wait for approval.

Phase 2 - after approval:
- copy a small test batch first
- verify checksums and openability
- keep an undo manifest
- stop on any collision or uncertainty

Codex and Sites: Build Original Tools With Acceptance Tests

Matt shows a reimagined classifieds site, an original puzzle game called Echo Garden, AI-assisted videos, Blender control, 3D city experiments, a FutureTools overhaul, BuseyBench, and a tiny macOS utility that toggles Dock visibility from a Stream Deck. The strongest examples are not the largest. They solve one specific problem for one operator.

Build [small original tool] for [specific user and job].

Before coding:
1. inspect the current workspace and constraints
2. propose the smallest useful scope
3. define functional, visual, mobile, accessibility, and failure tests
4. list external services, permissions, and data collected

During the build:
- use original content and assets
- verify each major path in the browser
- keep secrets outside source code
- do not publish or connect production accounts

Stop at a tested local preview and report the evidence, limitations, and next
approval needed.

How to Build Your Own Control Tower

Matt's private Wolf Control Tower combines AI-lab news, newsletter signals, brand mentions, website traffic, audience counts, and week-over-week change. It is a good example of Codex-native software: a private product that helps its owner manage a system instead of manually checking many dashboards.

Build it private-first. Start with three panels: what changed, what needs attention, and what can wait. Do not begin with every metric. A dashboard becomes useful when it reduces decisions, not when it reproduces every source in a prettier grid.

Route Sol, Terra, and Luna by Task

OpenAI positions Sol as the flagship for ambitious agentic work, Terra as the balanced model, and Luna as the fast affordable option. The exact best route depends on your plan, tool surface, reasoning level, and acceptance criteria. Start with the least expensive model that reliably passes the task.

Measure cost per accepted result: model usage plus operator review plus retries. A cheaper model that needs four corrections can cost more than a stronger model that finishes once. A stronger model used for every classification task is still wasteful.

A Safer Seven-Day Rollout

1. Day 1: install the official app and map Chat, Work, Codex, Sites, browser modes, and current plan limits.
2. Day 2: create one read-only project with one low-risk source and run the five-task acceptance test.
3. Day 3: create one editable voice profile from a narrow date range and test three draft types.
4. Day 4: run the one-page weekly review. Correct dates, ownership, and source labels.
5. Day 5: build one private mini app or local utility in Codex with explicit tests.
6. Day 6: create one scheduled summary with no write access and verify the first unattended run.
7. Day 7: review permissions, remove unused connections, record time saved and corrections, and decide whether one new source is justified.

Bottom Line

Matt Wolfe's guide is strongest as an operating-system demonstration. ChatGPT 5.6 can move between context, connected services, browser research, documents, code, Sites, schedules, and private software. The Wolf Control Tower shows where that leads: a personal system built around the operator's actual work rather than another generic SaaS dashboard.

The prompting lesson is equally important. Good prompts define a source boundary, an output, a verification method, and a stop condition. Good systems add one more thing: an approval gate before consequence. Start with one source, one result, and one week of evidence. Then expand because the workflow earned more access, not because the demo made total access look exciting.

Sources

• Matt Wolfe: Complete Guide to ChatGPT 5.6 + Prompting Guide - sponsored by OpenAI, as disclosed in the video.
• Matt Wolfe on YouTube, Matt on X, and FutureTools
• Official ChatGPT download
• OpenAI: ChatGPT for your most ambitious work
• OpenAI: GPT-5.6
• OpenAI Help: Built-in browser in the desktop app
• OpenAI Help: Cloud browser in ChatGPT
• OpenAI Help: Apps in ChatGPT
• OpenAI Help: Scheduled Tasks
• OpenAI Help: ChatGPT agent safety
• OpenAI: Codex for work
• JQ AI SYSTEMS: ChatGPT Work
• JQ AI SYSTEMS: ChatGPT 5.6 Hidden Features
• JQ AI SYSTEMS: Dan Shipper's GPT-5.6 Work OS