Plan a Project With AI: Turn a Goal Into a Task List That Holds Up

An AI project plan is easy to get and hard to trust. Ask any model to plan a launch and it hands back clean phases, plausible tasks and a timeline that happens to fit whatever deadline you named. The plan is not wrong so much as incomplete, and the gaps are invisible precisely because everything that is there looks reasonable.

Can AI actually write a project plan?

It can write most of one. A current model is strong at the parts of planning that are pattern work: naming the phases, listing the tasks a project of this type usually needs, spotting the obvious dependencies, and turning all of it into a table you can paste into a tracker. It is weak at the parts that depend on your reality: how long your team actually takes, who is already at capacity, which approval quietly costs three weeks because it lives on one person's calendar. So use the models for coverage and structure, and keep estimates and ownership as human decisions.

What goes in the brief before any model starts planning?

Five things, written once and reused for every model: the goal stated as a finished outcome, the hard deadline, the fixed constraints (budget, headcount, tooling), the definition of done, and what is explicitly out of scope. Anthropic's prompting guidance makes the same point in different words: a model has no context on your norms or your situation unless you state it, so state it. Draft that brief in Solo mode in aiDex, tighten it, then stop editing. Everything below reuses the same brief, which is what makes the plans comparable.

Why send the same brief to three models at once?

Because planning is a coverage problem, and one model gives you exactly one coverage pattern. Run the brief in Compare mode and GPT-5.4, Claude Opus 4.8 and Gemini 3.1 Pro each return an independent breakdown, side by side, with none of them seeing the others first. Read the three as a union rather than a contest. Tasks that appear in all three are table stakes. The task that appears in only one, a legal review, a data migration, a customer comms step, is the payoff: that is the item your own draft was also going to miss. Pick which models sit on the panel in the Dex. Use your own provider keys or the ones we manage, and pick the models you want.

How do I merge three breakdowns into one plan?

Move the three answers into Team mode and ask the models to argue about order, not about content. In Team the models share one thread, so the instruction is narrow: produce a single merged list, mark every dependency, and flag each task where you disagree about sequence. Those disagreement flags are the useful output. A sequence the models cannot settle is usually a dependency that is not settled in real life either, which means it is a question for a human before the work starts, not after.

How do I stress-test the merged plan?

Score it before you commit to it, against criteria you wrote down first. Run Judge mode over the merged plan with four: coverage (does any phase have zero tasks), sequencing (does anything start before its input exists), realism (is any single task quietly doing four jobs), and ownership (does every task name a role). Writing the criteria before you see the scores is what stops Judge from simply rewarding the longest plan. Then send the winner down a Pipeline, Draft, Critique, Revise, Polish, so what lands is a document you can hand to the team instead of a chat transcript. From there, a pre-mortem on the finished plan is the natural next pass.

What should I never let the plan decide?

Estimates, capacity and anything political. A model has no record of how long your team took last quarter, so any duration it offers is a genre convention rather than a forecast. Replace every number with your own history or your tracker's. The same goes for ownership: a model will cheerfully assign three parallel workstreams to a role held by one person who is already booked. Treat the output as the complete list of things that must happen and the rough order they must happen in, and treat the calendar as yours. If you are choosing between two ways to run the project rather than sequencing one, a decision matrix or a strategy table is the better tool.

Can I do this with a confidential project?

Yes, by running the panel locally. Point aiDex at your Ollama install and the brief, the breakdowns and the merged plan never leave your machine. You trade some quality against frontier models and you keep the identical five-step routine, because the value here comes from three independent passes rather than from any one model being brilliant. For mixed sensitivity, write the brief and merge locally, and send only a redacted version to cloud models.

How long does the whole routine take?

One sitting. The brief takes the longest, because it is the only step that requires you to decide anything; the three breakdowns run in parallel, the merge is one prompt and the scoring is one more. If it is dragging, the brief was too vague and the models are filling your gaps with genre. Save the panel in Teams so the next project starts from the same three models instead of a blank prompt. The wider pattern, and where the other modes fit, is laid out in multi-model AI workflows, and the mechanics of running models against each other are in how to compare AI models.