TL;DR
You have a goal like “launch v1 in 8 weeks” and every planning session ends with you doing something easier instead. The fix: have AI decompose the goal into milestone → week → day, one level at a time, with a measurable outcome at each level and every daily task capped at 2 hours. Do not let AI plan past week 2 — anything further out will change. Below are the three exact prompts to paste, the tool to use for each, and the one number that should change how you read every AI time estimate.
Why one level at a time
The mistake everyone makes is asking AI to plan the whole quarter on day one. You get a beautiful 32-task plan, and by week 2 half of those tasks are obsolete because week 1 taught you something. You stop trusting the plan, and an untrusted plan is just a guilt generator.
So you decompose exactly one level deep at a time: goal into four milestones, milestone 1 into weeks, week 1 into days. You re-decompose each subsequent week only after the previous one ships. This keeps the plan small enough to trust and recent enough to be right.
Who this is for
Founders shipping a v1, students preparing for a thesis or exam, side-hustlers stealing five hours a week from real life, and individual contributors staring at a quarterly OKR they cannot start. It also works for tech leads scoping a feature: the same milestone-week-day decomposition produces a Linear epic with sensible weekly tickets.
Pick the right tool for each step
You do not need a special “planning app.” The whole workflow runs in a general assistant, but the tool you pick changes how durable the plan is. As of June 2026:
| Tool | Plan / price | Why it fits goal planning | Watch out for |
|---|---|---|---|
| ChatGPT Projects (GPT-5.5) | Plus $20/mo | Persistent project space; ChatGPT Tasks can schedule the weekly re-decomposition as a recurring reminder | Free tier shows ads (US, since Feb 2026) and has tight limits; in-app context on Plus is ~320 pages, full 1M only on $200 Pro |
| Claude Projects (Sonnet 4.6 / Opus 4.7) | Pro $20/mo ($17 annual) | Searchable knowledge base — drop your retros and past plans in once, every chat sees them; 1M-token context standard | No built-in scheduler; you set the Friday reminder elsewhere |
| Gemini (3.1 Pro) | Google AI Pro $19.99/mo | 1M context plus native Google Calendar / Workspace reach for capacity-based planning | Was renamed from “Gemini Advanced” in early 2026 |
For most people: Claude Projects for the thinking (it holds your real velocity history across chats), with a ChatGPT Task doing the recurring Friday nudge. If you only want one tool, ChatGPT Projects covers both. ChatGPT Tasks is in beta on Plus, Pro, and Teams and caps you at 10 active tasks, which is plenty for one recurring planning reminder per goal.
Before you start
- Write the goal in one sentence with a deadline. “Launch v1 in 8 weeks” works. “Get serious about ML” does not.
- Name the definition of done. If the goal is “launch v1,” does that mean signups open, first paying customer, or App Store live?
- Block 30 to 45 minutes of uninterrupted time. Decomposition is not a 5-minute task.
- Know your real weekly capacity in hours. If you have 6 hours a week, AI will happily plan 20 — that is a fantasy plan.
- Have your tracker open (Linear, Notion, Things, Todoist). Each day-level task goes in immediately.
The three prompts
Paste these in order, one level at a time. They are deliberately strict about measurable outcomes, because that is where AI plans fall apart.
Prompt 1 — milestones.
Goal: [your goal in one sentence with a deadline].
Definition of done: [what "done" concretely means].
My real capacity: [N] hours per week.
Break this into exactly 4 milestones in sequence. Each milestone must end
with a MEASURABLE outcome, not a verb. Reject your own milestone if its
success criterion is "improve", "build", "set up", or "work on". Give each
milestone a target end-week. Do not plan inside the milestones yet.Prompt 2 — weeks (milestone 1 only).
Take milestone 1 only. Break it into weekly outcomes for [N] weeks.
One sentence per week, ending with exactly what is shippable, demoable,
or measurable on Friday. Assume [N] hours of capacity per week and no more.Prompt 3 — days (week 1 only).
Take week 1 only. Break it into daily tasks of 2 hours or less.
Mark dependencies between days (e.g. "blocked by Tue"). Add a Day 0
"setup" task for anything I must install, sign up for, or get a
credential for before real work starts. Reject any task over 2 hours
and split it.After each output, push back on anything that fails the measurable-outcome test. The single most useful follow-up is a constraint you carry: I do not have backend skills — re-decompose with that constraint. The plan gets dramatically more honest.
Step by step
- State the goal and time horizon in one sentence. Example:
Launch v1 of my note-taking app in 8 weeks with 50 beta users. - Run Prompt 1. Reject any milestone whose success criterion is a verb.
- Run Prompt 2 on milestone 1 only.
- Run Prompt 3 on week 1 only.
- Capture each daily task in your tracker with the milestone tag and the week tag. Tag dependencies so blocked tasks do not stare at you.
- Execute week 1. Re-run Prompt 2/3 for week 2 only after week 1 is done — what you learned in week 1 will invalidate week 3’s plan anyway.
- After milestone 1 ships, re-run Prompt 1’s logic for milestone 2 from scratch. Do not trust the plan AI made for milestone 2 eight weeks ago.
The one number that should change how you read estimates
AI is confident and consistently wrong about time, and so are you. In a 2025 randomized study by METR, experienced open-source developers predicted that using AI tools would cut their task time by 24%; after finishing, they believed it had saved 20%. The measured result was the opposite — AI tooling increased completion time by 19%, mostly because of review and validation overhead that nobody budgets for (METR, 2025).
The lesson for planning: do not trust raw AI time estimates, and do not trust your own gut either. Cut every daily task estimate by about 30% on entry, then track an “AI estimate vs actual” ratio per category over a few weeks. For coding tasks the real-world ratio is often 2 to 3x AI’s guess; for writing it is closer to 1.2x. Apply your own ratio to future estimates instead of arguing with the model.
First-run exercise
Pick a small goal first — one you can finish in two weeks, like “ship a personal site.” Run all three prompts. Then do not start executing for 24 hours; sleep on it. The next morning, look at day 1’s tasks: can you start the first task in 90 seconds without any further thinking? If yes, the decomposition is good. If you have to think about “where do I begin,” a day-level task is still too abstract — push it one level smaller.
Quality check
- Every milestone has a measurable outcome. “Improve UX” is not measurable. “First-time user can sign up and create a note in under 60 seconds” is.
- Every weekly outcome ends with something shippable, demoable, or measurable on Friday. Otherwise the week disappears into busywork.
- Every daily task is 2 hours or less. AI over-estimates how much fits in a day; cut every task by 30%.
- No task references a tool you have not installed or an API you have not signed up for. That belongs in the Day 0 setup task.
- Dependencies are explicit. If “deploy to prod” depends on “buy domain,” the dependency must be visible.
How to reuse this workflow
- Save Prompt 1, 2, and 3 as three snippets. Each is a one-line tweak per goal.
- Keep a
goals/folder with one file per quarterly goal. Drop it into a Claude Project’s knowledge base so every future chat sees your real velocity history. After four quarters you have a corpus of how long real things took versus AI’s estimates. - Set a recurring ChatGPT Task for “Re-decompose this week” every Friday, so the cadence does not depend on memory.
- Re-run the milestone decomposition every two weeks even when nothing has slipped. New information should change the plan.
- After a goal ships, write a one-paragraph retro and paste it into the next quarter’s planning prompt. AI plans better with examples of your real velocity.
Common mistakes
- Decomposing all four milestones into daily tasks at once. By week 2, half the day-level tasks are obsolete and you stop trusting the plan.
- Tasks larger than 2 hours. They sit in the tracker for days because starting feels too expensive.
- Milestones with no measurable outcome. “Improve onboarding” expands forever; “reduce signup-to-first-action time to under 60 seconds” finishes.
- Trusting AI’s time estimates without applying your personal velocity ratio.
- Skipping the Day 0 setup tasks. The first real task often blocks on installing a dependency, signing up for a service, or finding a credential.
- Replanning instead of executing. If you have decomposed the same week three times, the problem is not the plan.
FAQ
- Which AI should I actually use for this?: Any of the three in the table works. As of June 2026, use Claude Projects (Pro, $20/mo) when you want the plan to remember your past velocity across chats, or ChatGPT Projects (Plus, $20/mo) if you want the planning and the recurring Friday reminder in one place via ChatGPT Tasks. Gemini (Google AI Pro, $19.99/mo) is the pick if you plan around Google Calendar capacity.
- My goal keeps changing. Is this workflow still useful?: Yes — re-decompose more often. Treat the milestone level as stable for the quarter and re-decompose weeks every Friday.
- What if AI suggests a milestone I cannot do?: Push back in the conversation: “I do not have backend skills. Re-decompose with that constraint.” The plan gets dramatically more useful.
- Should daily tasks be 1 hour or 2 hours?: Two hours for focused work, one hour if the task is research-heavy or requires switching contexts. Smaller is always safer.
- How do I handle goals that depend on other people?: Mark those tasks
waiting on Xand keep a separate sub-plan for the dependency. Do not let waiting tasks block your own work. - Can I use this for personal goals like fitness?: Yes. The milestone-week-day structure works for any goal with a deadline. Replace “shippable” with “measurable” — pounds, miles, minutes.
- What if I miss week 1’s outcome?: Do not try to catch up; re-decompose. Ask AI to compress milestone 1 into three weeks instead of four, accepting that something has to drop.