5 AI Prompts to Debias Your Plans (Use Before You Commit)

The cognitive biases that distort plans are not failures of intelligence. Researchers Daniel Kahneman and Amos Tversky documented the planning fallacy in the 1970s — the systematic tendency to underestimate time, cost, and risk on projects — and subsequent decades of research have confirmed it applies equally to experienced professionals and novices. Bent Flyvbjerg’s analysis of large infrastructure projects found average cost overruns of 45%, and the pattern holds in smaller personal and professional planning contexts too. Knowing about a bias does not reliably protect you from it, because the bias operates before conscious deliberation enters the picture.

What AI adds to this problem is not immunity to bias — AI systems have their own distortions — but a structured adversarial perspective. When you run a pre-mortem or a steel-man challenge through an AI prompt, you’re forcing yourself to articulate specific failure modes before you’re emotionally invested in the outcome. The research on pre-mortems (Gary Klein’s work, later extended by others) suggests this sequence — generate failure explanations before commitment — produces more accurate forecasts than post-hoc analysis.

These five prompts are designed to be used before you finalize any significant plan. Each addresses a specific cluster of cognitive bias. Paste your plan summary where indicated and run the prompt as written.

Prompt 1: The Reference Class Check

Addresses: Planning fallacy, optimism bias

Use when: you have a timeline or budget estimate and want to pressure-test it against how similar projects typically go — before you’ve told anyone what the number is.

I'm planning [brief project description]. My current estimate is [timeline and/or budget].

Before I commit, compare this to the base rate for similar work:

1. What category does this project belong to?
2. What is the typical range of time and cost for this category, based on general patterns?
3. What are the three most common reasons this type of project takes longer than planned?
4. Given typical overrun patterns, what would a calibrated estimate look like?

What to do with the output: If the calibrated estimate is meaningfully higher than your current estimate, adjust before committing. If you still prefer the original estimate, you need to name specifically why your project is different from the reference class — and that reasoning should be explicit, not intuitive.

What good output looks like: The reference class is specific enough to be meaningful — “mid-size brand identity projects for B2B companies” rather than “creative projects.” The overrun reasons are concrete and recognizable (“scope expansion during stakeholder reviews,” not “things take longer than expected”). A sign of failure: the calibrated estimate is suspiciously close to your current estimate. The point of reference class forecasting is that it almost always suggests more time and cost than internal estimation does. If the output validates your estimate without adjustment, the reference class is probably too broad.

Prompt 2: The Pre-Mortem

Addresses: Confirmation bias, optimism bias, narrative fallacy

Use when: you’re about to commit to a plan and want to surface the most plausible failure modes before optimism bias makes them invisible. This is the most broadly applicable prompt in the set — run it on any significant plan.

I want to run a pre-mortem on this plan: [paste your plan or a one-paragraph summary].

Assume it is [target end date] and the plan has clearly failed—not a minor miss but an obvious failure.

Generate the five most plausible explanations for why it failed. For each:
- Describe the failure mode specifically
- Identify which assumption in the plan this failure invalidates
- Rate the probability as: high, medium, or low

Focus on common, realistic failure modes—not rare catastrophes.

What to do with the output: For each high-probability failure mode, decide: revise the plan to reduce the probability, add a contingency, or document a trigger condition that will prompt reassessment if it starts materializing.

What good output looks like: The failure modes are recognizable and specific to your plan — not generic risks that apply to any project of any kind. At least two or three should reference something you actually assumed rather than obvious external risks. A sign of failure: all five failure modes are low-probability tail risks (“key team member gets sick,” “major market shift”). These are real but not diagnostic. Push back and ask for the common, boring failures — the ones that explain 80% of missed plans.

Prompt 3: The Assumption Audit

Addresses: Narrative fallacy, overconfidence, Dunning-Kruger

Use when: your plan rests on a chain of interdependent claims and you want to know which of them are load-bearing and which have actually been verified.

Here is my plan: [paste plan].

Identify the key assumptions each major milestone depends on. For each assumption, categorize it as:

- Verified: backed by direct evidence or confirmed data
- Inferred: reasonable analogy from related experience
- Untested: assumed or believed but not yet confirmed

For each untested assumption in a critical-path milestone, describe what evidence would move it to verified.

What to do with the output: Untested assumptions in critical-path milestones are your highest-risk items. For each one, decide whether you can run a quick test or confirmation before committing the full plan — a short conversation, a prototype, a literature check.

What good output looks like: The audit surfaces assumptions you hadn’t consciously named — the ones embedded in phrases like “once the team is aligned” or “assuming normal market conditions.” A sign of failure: everything comes back as “verified” or “inferred” with few untested items. A plan with no untested assumptions is either unusually well-researched or the AI has been too charitable. Ask it to specifically look for assumptions that are taken for granted rather than confirmed.

Prompt 4: The Sunk Cost Trigger

Addresses: Sunk cost fallacy, status quo bias, present bias

Use when: you are starting a significant project and want to define in advance the conditions under which you would pause, pivot, or stop — before you’ve invested enough to feel emotionally committed to continuing.

For this plan, help me define pre-committed update triggers—conditions I can define now, before I am invested in the outcome.

Describe a specific, observable condition for each of these:

1. Timeline trigger: at what point does a delay indicate the original estimate was wrong rather than just delayed?
2. Assumption trigger: what evidence would indicate a core assumption has been invalidated?
3. Resource trigger: what signal shows the capacity estimates were significantly off?
4. Strategic trigger: what external change would make the rationale for this plan obsolete?

Make these concrete and observable, not vague.

What to do with the output: Write these triggers into your plan document or project notes. Review them at the start of each week. If a trigger condition is met, the response is not panic — it is a pre-planned reassessment conversation you already agreed to have.

What good output looks like: Each trigger is a sentence you could read to another person and they’d know unambiguously whether the condition has been met. “Two consecutive weeks where the timeline slips by more than three days” is a trigger. “If things feel significantly behind” is not. A sign of failure: vague triggers that require judgment to interpret — they’ll be rationalized away under the same sunk-cost pressure the prompt is designed to preempt.

Prompt 5: The Steel-Man Challenge

Addresses: Confirmation bias, availability heuristic, narrative fallacy

Use when: you’ve already decided on an approach and want to pressure-test whether you’ve considered the strongest version of the alternative before committing.

Here is my plan: [paste plan or summary].

Steelman the case against this plan. I want the strongest possible argument that:

1. The approach is wrong—there is a better way to achieve this goal
2. The timing is wrong—this plan should be done at a different time or in a different order
3. The assumptions are wrong—the key beliefs this plan rests on are likely incorrect

Do not soften the critique. I want the strongest version of each objection.

What to do with the output: A strong steelman that you cannot rebut is a signal worth taking seriously. You do not have to abandon your plan because an AI found objections — but if you cannot answer the objections, the plan has a problem you have not yet addressed.

What good output looks like: Each objection makes you slightly uncomfortable, because it engages your actual reasoning rather than generic alternatives. A sign of failure: the objections are polite and easily dismissed (“some might argue the timeline is ambitious”). The prompt explicitly asks for the strongest version of each objection — if the output is soft, tell the AI to remove the qualifications and state the objection as directly as possible.

Common Pitfalls

Pitfall: running these prompts after the plan is already shared or announced. Social commitment — having told stakeholders or a team what you’re doing — makes it significantly harder to act on debiasing output, because changing the plan now carries a social cost. Fix: treat these prompts as a pre-announcement gate, not a post-announcement review. Run them while the plan is still yours to change.

Pitfall: using vague plan descriptions as input. A one-sentence summary like “I’m planning a product launch” produces reference class data so broad it’s useless and pre-mortem failure modes so generic they could apply to anything. Fix: paste a real paragraph — the scope, the timeline, the key dependencies, and the assumptions you’re making. The more specific the input, the more specific and useful the output.

Pitfall: treating AI validation as actual validation. If you run the steel-man and the AI can’t find strong objections, that doesn’t mean the plan is sound — it may mean the AI is working from the same framing you’ve already accepted. Fix: after running these prompts, find one person who is skeptical of your plan (not someone invested in its success) and ask them to read the AI output and add anything it missed.

Pitfall: running all five prompts once and considering the plan debiased. The reference class and pre-mortem are most useful before commitment; the sunk cost triggers are only useful if you review them regularly. Fix: schedule the trigger review at the same cadence as your project check-ins. Debiasing is a process, not a one-time event.

Start here: Copy Prompt 2, paste a summary of your current plan, and run it. The pre-mortem addresses more biases simultaneously than any other single prompt in this set and takes about ten minutes.

For the complete framework — including how to sequence these prompts against a planning calendar and how to handle the output when it conflicts with stakeholder expectations — see the Complete Guide to Cognitive Bias in Planning.

Tags: AI-prompts, debiasing, cognitive-bias, planning, quick-win