How to Use ChatGPT and Claude to Plan Projects (and What Not to Delegate to AI)

You open ChatGPT or Claude, type "build the plan for a CRM implementation project," and in ten seconds you get a ten-phase plan with nice names. It looks ready. Except it doesn't know your team has two people, that the client only approves on Thursdays, or that the last implementation got stuck on data migration. The plan is plausible, and plausible isn't the same as executable.

AI plans well when you give it context and use the result as a draft. This article offers a four-step prompt roadmap (scope, tasks, risks, and schedule), what you shouldn't delegate, and how to bring the result into your management tool without retyping everything.

Before the prompts: give context

The difference between a generic plan and a useful one is in the first paragraphs you write. Before asking for anything, paste or describe:

• the project's goal and who the client or requesting department is;
• the final deadline and what happens if it's missed;
• who works on the project, with roles and availability (for example, "two people, 20 hours a week each");
• known constraints: budget, mandatory tools, third-party dependencies;
• what went wrong in similar projects.

If there's a briefing, a proposal, or meeting minutes, paste the text. The model works better transforming real material than inventing from a single sentence.

Step 1: scope

The first use of AI is to find out what's still undefined. Instead of asking for a finished scope, ask for questions.

I'm going to give you a project briefing. Before writing any plan, list the ten questions you would ask the client to pin down the scope, in order of impact. Also point out what seems to be out of scope and deserves confirmation. Briefing: [paste here]

Answer the questions, and only then ask for the scope.

Based on my answers, write the project scope in three blocks: what's in, what's out, and the assumptions. Each "in" item should have a verifiable acceptance criterion.

The acceptance criteria are the part that pays off the most. An item like "reports delivered" becomes "three reports (sales, inventory, finance) opened without errors by two of the client's people," and that's what later supports the acceptance form.

Step 2: breaking into tasks

With the scope approved by you, ask for the work structure.

Break the scope above into deliverables and, for each deliverable, into tasks of at most two days of work. For each task, give: suggested responsible role, estimate in hours, dependency on other tasks, and definition of done. Deliver it as a table.

Then ask for a critical review of what it wrote itself:

Reread the list and point out tasks that are too big, missing tasks (testing, approval, client communication, training), and dependencies that don't make sense.

This second pass catches a good share of the oversights. Even so, read the whole list: AI tends to forget the "invisible" tasks, like waiting for third-party approval, and underestimates integration ones.

Step 3: risks

Here the model helps as an ego-free reviewer, as long as you feed it the history.

Consider this plan and the team's context. List the twelve biggest risks, each with: what could happen, early warning sign, probability (low, medium, or high), impact, and a prevention action. Prioritize risks tied to client dependency, lack of decisions, and scope change.

Of the twelve, pick three to five that are real for your case and turn each into a follow-up task, with an owner. A risk that stays only on the list is decoration.

Step 4: schedule

Ask for a sequence, not final dates.

Organize the tasks into an execution order, indicating what can run in parallel and the critical path. Assume the team has two people with 20 hours a week each. Show where the load exceeds capacity.

The value is in seeing the order and the capacity bottleneck. The dates are for you to adjust. Anyone following the plan in a schedule with dependencies and critical path can check whether the suggested order holds up once the real dates come in.

What not to delegate to AI

Decision	Why it stays with you
Political priority	Choosing what to do first is deciding who will wait. It depends on relationships, history, and interests the model can't see.
Deadline promised to the client	AI produces a plausible estimate; you sign a promise. The basis should be the team's real history, not a generated number.
Staffing	Who does what involves vacations, friction, learning curves, and conversations that aren't in any text.
Accepting or refusing a scope change	It's a commercial and relationship negotiation.
Confidential information	Personal data and contracts should be handled according to your company's policy before pasting into any tool.

The practical rule: use AI to widen your options and to review, and keep for yourself everything that involves a commitment to another person.

How to bring the plan into the tool

There are three paths, from the most manual to the most integrated.

1. Copy and import. Ask the model for the list as a table, review it, and import it as tasks. It's the simplest path and works in any tool that accepts imports.
2. Inside the tool. In Tasskee, the AI assistant works with the project's own data, in planning, execution, and automatic modes, with a record of what it did. It uses your provider's key (OpenAI, Gemini, or Claude), with no credits and no markup on usage. The math is in AI without credits, with your own key.
3. Your agent operates the tool. Tasskee's MCP server lets Claude or ChatGPT itself, in the conversation where you did the planning, create the tasks directly in the project. You plan by chatting and the agent writes them in, with no copy and paste. The details are in how AI works inside the project. The feature is on the Pro plan.

On any path, the same discipline applies: review before saving. Forty tasks created at once don't save time if you have to delete ten.

Signs the generated plan is no good

• All phases have similar durations. A real project has short and long stages.
• There isn't a single task for waiting, approval, or a client decision.
• The estimates are all round numbers with no relation to the team's capacity.
• The risks look like those of any project in the world. If they fit everyone, they don't fit yours.

Two prompts for the next day

Planning doesn't end on day one. Two questions help keep the plan alive when you paste the project's current state into the conversation.

Here is the project's current state: what was completed, what's late, and what changed since the original plan. Point out the three decisions I need to make this week and the consequences of each alternative.

Write a status summary for the client in five lines: what was delivered, what's in progress, what depends on them, the main risk, and the next important date. Direct tone, no jargon.

In both cases, AI organizes and drafts; you check the facts and decide what's said to the client.

A one-hour roadmap

1. Ten minutes writing the context.
2. Ten minutes on the scope questions and answers.
3. Fifteen minutes on the task breakdown and critical review.
4. Ten minutes on risks, choosing the ones that become tasks.
5. Fifteen minutes on the schedule and the capacity check, adjusting dates with what you know about the team.

One hour, instead of a day of planning, with the advantage that the decisions that count are still yours. If the result ends up in a single place where tasks, schedule, and goals live together, AI and the team end up working on the same plan.