Most project plans fail the same three ways: they list tasks without ranking by impact, they never name a single accountable owner, and they bury risks in a slide nobody reopens. The 11 prompts below force milestone-level deliverables, explicit assumptions, a risk register sorted by probability × impact, and a scope-cut menu prepared in advance — so when reality hits in week three, you cut on purpose instead of slipping in silence.
These are model-agnostic. Pair them with project risk analysis prompts for deeper risk work.
TL;DR
- Run the prompts in order: goal → milestones (#1) → tasks (#2) → sprint (#3), then layer risk (#4), ownership (#5), and critical path (#6) on top.
- Paste your real backlog, dates, and team list into the
{...}placeholders. Vague inputs produce generic plans. - For plans that need you to paste a lot of context (#3, #4, #6, #8, #9), use a 1M-token model: Claude Opus 4.7 / Sonnet 4.6, Gemini 3.1 Pro, or ChatGPT Pro ($200). Free and Plus tiers truncate long pastes.
- The highest-leverage prompts are #4 (risk register), #8 (scope-cut menu), and #9 (pre-mortem). Skip those and you plan for the happy path only.
Which model should you run these on?
These prompts are plain text and work on any chat model, but two things matter for planning: how much context you can paste, and how well the model keeps it straight. As of June 2026:
| Model | Standard context | Plan / price | Best for |
|---|---|---|---|
| Claude Sonnet 4.6 | 1M tokens | Claude Pro $20/mo | Daily driver; structured, reliable output |
| Claude Opus 4.7 | 1M tokens | Claude Max $100/$200 | Big multi-team programs, deepest reasoning |
| Gemini 3.1 Pro | 1M tokens | Google AI Pro $19.99/mo | Largest pastes, Workspace docs |
| ChatGPT (GPT-5.5) | ~320 pages in-app on Plus | Plus $20 / Pro $200 | Quick plans; full 1M only on $200 Pro |
Practical rule: a one-page goal-to-milestone plan (#1, #2) runs fine on any free tier. Once you paste a full backlog or an existing plan (#3, #4, #6, #8, #9), move to a 1M-token model so nothing gets truncated. Note: ChatGPT’s in-app context on the $20 Plus plan is roughly 320 pages; the full 1M-token window is reserved for the $200 Pro tier.
Best for
- Sprint planning
- New initiative scoping
- Multi-team programs
- Cross-functional launch plans
- Personal big projects (book, move, exam prep)
1. Goal → milestones (with assumptions)
My goal: {1 sentence}. Constraints: {time, team, budget}. Break into 4-6 milestones. For each: deliverable, success criteria, 1 key assumption that could invalidate it. Flag the milestone whose assumption is shakiest — that's where to test first.2. Milestone → tasks
Milestone: {description}. Owner: {team}. Available time: {N weeks}. Break into tasks of 2 days or less each. Mark dependencies between tasks. Estimate optimistic / realistic / pessimistic days for each.3. 2-week sprint plan
Given backlog (paste), team capacity {N points}, top priority {goal}: plan a 2-week sprint. Output: sprint goal in 1 sentence, committed items, stretch items, risks. Reserve 20% capacity for unplanned work.
{paste}4. Risk register
My project plan (paste). Identify top 8 risks. For each: probability (L/M/H), impact (L/M/H), mitigation, contingency, owner. Sort by probability x impact. Distinguish risks (uncertain) from issues (already happening).
{paste}5. RACI matrix
Project: {summary}. Stakeholders: {list with roles}. Build a RACI matrix for the 8 most important workstreams. Flag any workstream with more than 1 Accountable — that's a problem. Also flag any with zero Accountable.The hard rule for RACI is exactly one Accountable per workstream — never two, never zero. If you can’t pick one owner, the workstream is still too coarse; split it. Keep the Consulted column short, because every Consulted name you add is one more person whose sign-off can stall the work.
6. Critical-path identification
Below are my project tasks with dependencies. Identify the critical path. Mark the 3 tasks where slipping by 1 day slips the whole project by 1 day. Suggest 2 ways to compress the critical path: fast-track (parallelize) or crash (add resources).
{paste}7. Status update from plan + reality
My plan: {paste}. What actually happened this week: {paste reality}. Write a 150-word status update: on track / at risk / blocked, with one specific next-week priority and a single "what I need from you" ask.8. Scope-cut menu
I have 3 weeks left and likely 5 weeks of work. Plan: {paste}. Produce a "scope-cut menu": 5 items I could cut, what we lose, what we keep, who would notice. Rank by least painful first.9. Pre-mortem on the plan
Project: {summary}. Plan: {paste}. Imagine it's 6 months from now and the project failed. Write the postmortem: top 5 reasons it failed, sorted by likelihood. For each, name the leading indicator we'd see in week 2 if we were heading toward that failure.
{paste}The pre-mortem is the single most useful prompt here. The technique comes from research psychologist Gary Klein (HBR, 2007): imagining a failure has already happened, rather than that it might, makes people noticeably better at naming the real causes, because it sidesteps the optimism that kills honest risk talk during planning.
10. Stakeholder communication cadence
Project: {summary}. Stakeholders: {list with seniority + interest level}. Design a communication cadence: who gets updates, in what format (Slack thread / weekly email / monthly review), at what frequency, and what decision each update is meant to enable. Avoid status-for-status's-sake.11. Kickoff meeting agenda
Project: {summary}. Team: {list}. Design a 60-minute kickoff agenda that ends with these artifacts: a shared goal in 1 sentence, top 3 risks, owner per workstream, the 1 question we'd most regret not asking. Include rough timeboxes per section.How to drive these prompts well
- Replace every
{...}with real text. “Constraints: 6 weeks, 3 engineers, no external budget” gets a usable plan; “constraints: some” gets filler. - Chain them in one conversation so the model carries context: run #1, then feed its milestones into #2, then #2’s tasks into #6.
- When output is too generic, push back: “These risks are obvious. Give me 3 that are specific to a two-team handoff under a fixed launch date.”
- Treat AI estimates as a starting draft, not a commitment. Always re-check dates, owners, and any number before you publish the plan.
Common mistakes
- Listing tasks without milestones or outcomes — the plan reads like a to-do list, not a project.
- Capturing risks once at kickoff and never revisiting them as new ones emerge.
- No explicit Accountable per workstream — everyone assumes someone else owns it.
- Never identifying the critical path, so the team optimizes the wrong tasks.
- No scope-cut menu prepared in advance — when slip happens, cuts get made in a panic.
- Status updates that report activity (“did 12 things”) instead of outcomes against the goal.
FAQ
Which AI model is best for project planning? For most people, Claude Sonnet 4.6 on Claude Pro ($20/mo) is the sweet spot: a 1M-token context and reliably structured output. For very large multi-team programs, step up to Claude Opus 4.7. If you need to paste the most context or work inside Google Workspace, Gemini 3.1 Pro ($19.99/mo) matches the 1M window. (Figures as of June 2026.)
Can I run these on a free plan? Yes for the short prompts (#1, #2, #5, #10, #11). The paste-heavy ones (#3, #4, #6, #8, #9) can exceed free-tier limits and may truncate your input. If a plan looks like it ignored half your backlog, your paste was likely cut off — switch to a paid 1M-token plan.
Will the AI’s estimates be accurate? No, and you shouldn’t treat them that way. The optimistic / realistic / pessimistic estimates in #2 are a fast first draft to react to, not a schedule to commit. Have the people doing the work confirm every estimate and owner before the plan goes out.
What’s the difference between a risk and an issue? A risk is uncertain (it might happen); an issue is already happening. Prompt #4 keeps them in separate lists on purpose — mixing them buries live problems among hypotheticals, which is exactly how an active blocker gets ignored.
Do these work in Cursor or other coding tools? Yes. They’re plain text, so paste them into any chat model — including Cursor or Claude Code — when you want a plan for a software project alongside the code it touches.