_edited.png){kind=small}
How to Use Goals in ChatGPT Work With GPT-5.6 Sol
- Ben Angel
- 2 minutes ago
- 13 min read
Most entrepreneurs eventually reach a strange point with AI. The plans become smarter, but the amount of work they personally manage barely changes.
They ask ChatGPT for a strategy, receive twelve intelligent steps, feel a brief surge of optimism—and then become the project manager responsible for pushing every step across the finish line.
By the end, they have a polished to-do list wearing an AI costume—and a new unpaid job managing it.
That's exactly the problem Goals in ChatGPT Work are designed to solve.
Instead of asking ChatGPT for the next answer, you give it an outcome to achieve, boundaries to respect and a definition of done. Rather than stopping after the first response, ChatGPT Work can continue working toward that outcome—using your files, tools and instructions—before returning something you can review.
The simplest analogy is this:
A prompt is asking someone for directions. A Goal is handing them the destination, the car keys and asking them to call you when they arrive.
Human approval still matters. Goal mode simply gives ChatGPT room to handle the connected stages that can be delegated as one coherent assignment.
Quick Answer: What Are Goals in ChatGPT Work?
A Goal in ChatGPT Work is a persistent objective for a multi-step task. You start one by entering `/goal`, then describe the outcome, constraints and evidence that will prove the work is complete. The goal becomes both ChatGPT’s first instruction and its completion criteria. While it works, you can monitor progress, provide more context, change direction, pause the task or resume it later. Starting a Goal does not expand ChatGPT’s permissions or remove approval requirements.
If you want to move beyond isolated AI tricks and build practical systems like this into your business, my 28-Day AI Mastery Course provides the structured implementation path. It is designed to make AI useful when real work has to get finished.
In This Article
What Is ChatGPT Work?
ChatGPT Work is the part of ChatGPT designed for delegation rather than quick conversation.
Use ordinary Chat when you want an explanation, brainstorm, short rewrite or immediate answer.
Use Work when you want ChatGPT to complete a substantial task with a clear outcome: research a decision, analyze files, create a presentation, prepare a spreadsheet, assemble a campaign or manage a recurring workflow.
OpenAI describes Work as an agent that can gather context from your files and plugins, take action across workflows, and produce reviewable documents, presentations, spreadsheets, Sites and other finished work. You can follow its progress, answer questions and approve important actions while it works.
Think of Chat as calling an adviser for an answer. Work is bringing someone into the office, showing them the source material and asking them to return with the finished brief. In the AI-supported workflows I use, the bottleneck is rarely the first draft. It is checking sources, repairing weak sections and deciding whether the work is actually finished.
That is the gap Work and Goals are designed to close.
What Makes a Goal Different From a Prompt?
A prompt requests an output. A Goal defines an outcome and keeps ChatGPT working toward it.
If you type, “Give me a marketing plan for my new product,” ChatGPT will probably give you a marketing plan. It may be useful. It may even be excellent. But the assignment ends when the answer appears.
A Goal can be much broader: build the launch plan using supplied research, create the calendar and messaging, verify every claim, and stop before publishing.
The first request asks for information. The second delegates a result.
Here is the practical difference:
Output: A prompt usually produces one response. A Goal can pursue a multi-step outcome.
Depth: A prompt often stops after the first draft. A Goal can research, create, inspect and revise.
Next steps: A prompt depends on you to request every next step. A Goal uses the definition of done to choose the next step.
Working style: A prompt is optimized for immediate interaction. A Goal is designed for longer-running work.
Completion: A prompt may leave completion subjective. A Goal includes verification and stopping criteria.
This is why Goals are not simply “longer prompts.” A long prompt can still produce one answer. A Goal gives ChatGPT a finish line it can continue working toward.
A Goal defines the destination; an AI loop provides a repeatable process for checking and improving the work before ChatGPT declares it finished.
It is also different from an automation. A Goal pursues an objective inside a task. A scheduled task repeats, monitors or refreshes work at a particular time or when an event occurs.
You might use a Goal to perfect a weekly competitor report, then schedule that proven workflow once the result is reliable. OpenAI recommends refining recurring work in a normal task before scheduling it.
Why GPT-5.6 Sol Matters
GPT-5.6 Sol is OpenAI’s flagship model for complex, open-ended work involving research, computer use, analysis, marketing, coding and high-value deliverables. OpenAI recommends Sol when judgment, depth and polish matter more than raw speed or the lowest usage cost.
In plain English, Sol is the model you call when the assignment has moving parts. You would not hire an architect to change a lightbulb, or ask the cheapest handyman to design a twelve-storey building. Sol is intended for the second category.
OpenAI currently positions its GPT-5.6 family like this:
Sol: best for complex, open-ended or high-value assignments requiring greater depth and polish.
Terra: the practical everyday model balancing capability and cost.
Luna: the faster, more affordable option for clear, repeatable and higher-volume work.
For “research three markets, compare five competitors and draft the launch assets,” Sol is sensible. For “classify these survey responses into five predefined categories,” Luna may be more economical. A capable model still cannot rescue a vague objective or unreliable sources. Giving Sol a bad Goal is like giving a Formula 1 car directions that say, “Drive somewhere useful.”
When Should Entrepreneurs Use Goal Mode?
Use Goal mode when the work has a finish line but reaching it requires several connected steps.
Good candidates usually have four characteristics:
The task would take you meaningful time to complete manually.
It depends on several files, sources, tools or decisions.
It should produce something you can review and reuse.
Quality can be checked against specific criteria.
Examples include auditing a website, researching a market, turning customer interviews into a campaign, reviewing email performance, or converting several documents into a presentation and executive brief.
This is closely related to the shift I explored in AI Agents for Entrepreneurs: the leverage appears when AI owns a repeatable workflow, not when it produces another isolated caption or summary.
Do not use a Goal for every minor task. Asking Sol to spend an hour pursuing “rewrite this paragraph” is like sending a removal truck to pick up a sandwich.
Use ordinary Chat for quick answers and lightweight drafts. Use Terra or Luna for clearer, repeatable work. Save Sol-powered Goals for assignments where better reasoning and sustained execution could materially improve the outcome.
How to Use Goals in ChatGPT Work: Start Your First Goal
Starting a Goal is simple. Defining a useful one requires a little more thought.
1. Open ChatGPT Work
Select Work in ChatGPT. The desktop app can also use local files, apps and the browser when those tools are available.
2. Select GPT-5.6 Sol
Use the model and reasoning control beneath the composer. OpenAI’s current default Power setting uses GPT-5.6 Sol with medium reasoning. Increase the reasoning level for difficult tasks with several sources or trade-offs; lower it when the assignment is tightly defined.
3. Type /goal
Enter /goal in the composer. Goal mode is also available through Codex CLI and the IDE extension, although non-technical entrepreneurs will generally find the desktop app easier.
4. Describe the finished outcome
Do not merely describe the activity.
Weak: “Research my competitors.”
Stronger: “Create a competitor-positioning report for my AI course. Compare the five companies in the attached list across audience, promise, pricing, proof and sales mechanism. Identify gaps I can credibly own. Deliver a concise comparison, recommendation and links to current sources.”
5. Add your boundaries
Tell ChatGPT what it must preserve, avoid or ask permission before doing.
For example: use primary sources for prices, separate facts from interpretation, flag missing evidence, preserve the approved offer, and stop before contacting anyone or publishing anything.
6. Define what “done” means
Give ChatGPT a checklist it can use to inspect its own work.
The Goal is complete when all five competitors have been checked, every price has a dated source, the comparison table is filled, three defensible positioning opportunities are ranked, and unresolved questions are listed for my review.
Now ChatGPT knows when to stop polishing and when to keep working.
The Five-Part Goal Formula
OpenAI’s documentation emphasizes outcome, constraints and verification. For entrepreneurs, I would expand that into a five-part formula:
1. Outcome
What useful result should exist when the work is finished?
2. Context
Which files, plugins, customer data or examples should inform it?
The quality of a Goal still depends on the context behind it. If ChatGPT repeatedly lacks your standards, examples and business knowledge, start by building an AI brain for your business.
3. Constraints
What must remain unchanged, and which actions are forbidden?
4. Verification
What criteria prove the output is usable?
5. Approval boundary
Where must ChatGPT stop and involve you?
Use this reusable template:
Goal: Create [finished outcome] for [audience or business purpose]. Context: Use [files, sources, plugins, examples or prior work]. Constraints: Preserve [important facts, style or boundaries]. Do not [prohibited actions]. Verification: Check that [measurable completion criteria]. Flag anything you cannot verify. Approval: Stop before [sending, publishing, purchasing, deleting or changing important external information]. Return the completed draft for my review.
That is the difference between saying “help me with my launch” and handing over a properly defined assignment.
Practical Goal Examples for Entrepreneurs
Build an evidence-based email campaign
Review the attached customer interviews, sales-page copy and previous email results. Identify the beliefs preventing non-buyers from enrolling in x program. Build a five-email campaign that first reframes the problem, then demonstrates the limitations of the reader’s current approach, and finally presents the course as the implementation path. Preserve my conversational voice. Do not invent testimonials or performance claims. The Goal is complete when all five emails, subject lines, preview text and CTA copy are ready for review.
Audit the bottlenecks in a one-person business
Review my task list, calendar export and workflow notes. Identify repeated work that creates delay, lost revenue or decision fatigue. Rank the five best automation opportunities using frequency, attention cost and commercial impact. Recommend what to automate, what to delegate and what should remain human. Do not implement any automation. Return a decision-ready report with assumptions and next steps.
This complements my guide to AI automation for small business, where I explain why automating visible busywork can leave the real bottleneck untouched.
Create a launch package
Use the attached product brief, audience research and approved brand examples to create a launch package containing the positioning statement, campaign calendar, landing-page outline, five-email sequence, ten social posts and launch-day checklist. Keep all dates and prices unchanged. Cite factual claims to the supplied sources. Stop before publishing or scheduling anything. Verify that every asset uses the same core promise and CTA.
Prepare a business decision
Research the three CRM platforms on my shortlist using current primary sources. Compare total cost, setup time, email functionality, integrations, automation limits and data-export options for a one-person business with 12,000 subscribers. Identify unknowns and contract risks. Deliver a recommendation, but do not start a trial or contact sales.
Notice the pattern: the Goal gives ChatGPT a job that produces a reviewable business result. It does not ask the model to “do everything.” It narrows the assignment enough to make autonomy useful.
How to Steer a Goal While It Is Running
A Goal is persistent, not untouchable.
In the desktop app, a progress row appears above the composer. You can pause, resume, edit or clear the Goal. You can also send follow-up information while it runs, correct an assumption or add a missing constraint. OpenAI recommends using a side conversation when you want a progress recap or explanation without interrupting the main task.
Useful steering messages include:
“Keep the research, but make the recommendation less technical.”
“Do not change the approved course price.”
“Add this customer interview to the evidence before finalizing.”
“Pause after the audit and show me the proposed campaign structure.”
“The conclusion is too generic. Use the specific commercial risks identified in the source material.”
This is closer to managing a capable employee than operating a vending machine. Review the work, correct the direction and improve the brief. If you expect to lose connectivity, pause the Goal and resume it later.
Where Goal Mode Can Go Wrong
Goals make ChatGPT more persistent. They do not automatically make it more truthful, commercially aware or aligned with your judgment.
The main failure modes are predictable:
The Goal is too broad. “Grow my business” is not a Goal. It is a wish with Wi-Fi. Choose one reviewable outcome.
The sources are weak. Outdated screenshots and incomplete analytics can produce a beautifully organized mistake. Name authoritative sources and require uncertainty to be flagged.
There is no definition of done. The task may stop early, over-polish trivial details or return something that looks finished but cannot be used.
You remove the human approval point
A Goal does not grant broader access. It still operates under the same sandbox and approval policies and should pause when a decision is required.
Keep human approval before:
Sending emails or messages
Publishing content
Making purchases
Deleting information
Changing live customer or financial data
Making commitments to another person
The goal is not to remove yourself from judgment. It is to prevent your judgment from being buried under preparation work.
What Goals Cost and Who Can Use Them
ChatGPT Work and Codex share the same pricing, credits and usage limits.
OpenAI currently lists individual plans including Free, Go, Plus and Pro, alongside Business and Enterprise options. Plus is listed at $20 per month, Pro starts at $100 per month, and Business is listed at $20 per user per month when billed annually or $25 monthly, with at least two users. Pricing and availability can change, so check OpenAI’s current page before purchasing.
GPT-5.6 Sol is available in the ChatGPT desktop app, ChatGPT web, Codex CLI and the IDE extension, and it can use ChatGPT credits. OpenAI lists Sol as the most capable—and therefore more usage-intensive—member of the family. GPT-5.6 usage averages approximately 5–40 credits per message, but actual consumption varies with the model, context, reasoning, retrieval and tool use.
Multi-hour work involving files, research and tools will naturally consume more than a quick response. The useful question is not, “How many prompts did this use?” It is, “What finished work did those credits replace?” If you use Sol to alphabetize a list, you have hired the expensive architect to move the lightbulb again.
Inside my 28-Day AI Mastery Course, this is the capability I want entrepreneurs to build: choosing the right AI approach for the business outcome, not simply reaching for the most powerful model every time.
Goals Are the Missing Layer Between Prompting and Delegation
The most important shift is not from shorter prompts to longer prompts. It is from requesting outputs to defining outcomes.
Prompts still matter. But if every project requires you to request the next step and manually decide when the assignment is complete, you have become AI’s most overqualified project manager.
Goals give ChatGPT Work a durable objective, a finish line and permission to keep moving within the boundaries you set. GPT-5.6 Sol provides the depth for difficult assignments. Your role is to define the right outcome, provide credible context and remain responsible for the decisions that affect your customers, reputation and money.
That is the model I explore more broadly in The Wolf Is at the Door: AI is becoming more capable, but the advantage belongs to people who upgrade how they direct, evaluate and apply it.
Start with one finished result you have been postponing because it contains too many steps. Give ChatGPT the outcome, context, constraints, verification and approval boundary. You may discover that the limitation was never your prompts. It was that you kept giving AI errands when you could have given it responsibility.
Frequently Asked Questions
What is Goal mode in ChatGPT Work?
Goal mode gives ChatGPT a persistent objective for a multi-step task. The Goal becomes both the starting instruction and the criteria ChatGPT uses to decide whether the assignment is complete. You can monitor, steer, pause, resume, edit or clear the Goal while it runs.
How do I start a Goal?
Enter /goal in the ChatGPT desktop app then describe the outcome, constraints and verification criteria. If the outcome is unclear, OpenAI recommends using /plan first to clarify the approach and turn it into a measurable Goal.
What is the difference between a Goal and a prompt?
A prompt normally asks for one answer or output. A Goal defines a broader outcome that may require several steps, sources, tools and revisions. The Goal persists while ChatGPT works toward the stated definition of done.
What is the difference between a Goal and a scheduled task?
A Goal pursues a defined outcome inside a task. A scheduled task runs work once, repeatedly, when an event occurs or while monitoring for changes. Build and refine the workflow as a Goal or normal Work task before scheduling it when reliability matters.
Is GPT-5.6 Sol required for Goals?
No. Goal mode and model selection are separate. Sol is best suited to complex, open-ended and high-value work. Terra is a more economical everyday option, while Luna is intended for faster, clearer and repeatable tasks.
Can a Goal run without me watching it?
A Goal can continue working for an extended period, but it may pause when it needs information, approval or access. You can monitor progress and return to the task. For local work, your computer and required applications must remain available unless the workflow uses an appropriate remote or scheduled capability.
Can Goals send emails or publish content automatically?
Starting a Goal does not broaden ChatGPT’s permissions. Sending, publishing and other consequential actions remain subject to the available tools, sandbox and approval policies. For business use, explicitly tell ChatGPT to prepare drafts and stop for approval before any external action.
Can I change a Goal after it starts?
Yes. The desktop app allows you to pause, resume, edit or clear a Goal. You can also send follow-up messages to add context, correct direction or introduce a new constraint while the work is running.
Can I run more than one Goal at a time?
Separate tasks can run concurrently, and each task keeps its own context, messages, results and Goal. Avoid allowing two Goals to modify the same files or systems simultaneously because their changes may conflict.
Do Goals make ChatGPT’s work accurate?
No. Goals improve persistence and completion discipline; they do not guarantee factual accuracy. Use trustworthy sources, define verification criteria, require citations for important claims and review consequential outputs yourself.
Are Goals safe for confidential business work?
Safety depends on your plan, workspace policies, connected tools and the sensitivity of the information. Use the minimum necessary data, confirm your organization’s privacy requirements, restrict source and tool access, and preserve approval steps for consequential actions. Business plans include additional workspace and data controls, but you should still review current OpenAI policies for your situation.