The machine that keeps the receipts — what AI was claimed to do, and what it actually did.
Playbook · playbook

"Pre-mortem your plan: make AI find the holes before reality does"

· filed from inside the model

Turn a plan you're about to commit to into a ranked list of the ways it actually fails — and the cheapest thing to do about each one — before you spend the money.

LevelBeginnerTime10 minutesCost~$0.03ToolsAny chat LLM (Claude, ChatGPT, Gemini)Verified2026-06-17

The problem

You have a plan. A launch, a migration, a hire, a refactor, a trip you're quoting a client. It looks fine. It always looks fine — that's the problem. You wrote it, so it's built out of your assumptions, and your assumptions are exactly the things you can't see.

The failure isn't that you ignored the risks. It's that the real risk was never on your list. Three weeks later something obvious-in-hindsight tips over, and you say the thing everyone says: we should have seen that coming.

A pre-mortem fixes this. You imagine the plan has already failed and work backwards to why. It's a known debiasing technique (Gary Klein's, originally), and it works because "list the risks" invites optimism while "it's six months later and this was a disaster — what happened?" invites honesty. AI is unreasonably good at it: it has read ten thousand post-mortems of plans like yours and it has no ego invested in yours surviving.

The catch: if you prompt it lazily, it hands you a horoscope — "ensure clear communication," "monitor your budget," "be mindful of scope creep." True, useless, and you knew it already. The recipe below is the difference between that and a list you actually act on.

When to use this — and when not to

Use it when:

  • The decision is hard to reverse or expensive to get wrong (spending money, shipping to users, telling a client a number).
  • You've already got a plan written down. This sharpens a plan; it doesn't invent one.
  • You have ~10 minutes before you commit. The whole point is doing it before.

Don't bother when:

  • The action is cheap and reversible. If you can undo it in an afternoon, just do it — running a pre-mortem is slower than failing.
  • You haven't actually decided anything yet. Pre-mortem a vague intention and you get vague risks. Write the plan first.
  • The honest answer is you've already decided and you want cover. AI will give you a tidy risk list you can wave around, and you'll have learned nothing. That's not this tool's fault, but it's a real way people use it.

The recipe

Paste your plan in place of the bracketed part. The constraints in the prompt are doing the real work — the ranking, the specificity rules, the ban on generic advice — so don't trim them.

You are a skeptical operator who has watched plans like this one fail
many times. I'm about to commit to the plan below. Run a pre-mortem.

THE PLAN:
[paste your plan, including the goal, the timeline, the resources,
and any assumptions you're aware you're making]

Do this:

1. Assume it's [pick a date past the deadline] and this plan FAILED —
   not catastrophically, just the ordinary way things fail. Generate
   8-10 distinct failure stories. Each one is a specific causal chain:
   "X happened, which meant Y, so the plan died at Z." No vague risks.

2. Kill every failure mode that is generic enough to apply to any plan
   ("scope creep," "poor communication," "ran out of budget"). Replace
   it with the specific version that applies to THIS plan, or drop it.

3. Rank the survivors in a table by: Likelihood (H/M/L), Impact (H/M/L),
   and how early I'd see it coming (Early/Late). Late + High-impact
   failures go to the top — those are the ones that ambush you.

4. For the top 3, give me the cheapest possible action I could take
   THIS WEEK that would either prevent it or surface it early. Not
   "monitor X" — a concrete thing I do once.

5. End with the single assumption this whole plan rests on that, if
   wrong, makes everything else irrelevant. State it as a yes/no
   question I could go answer.

Be concrete and a little harsh. I'd rather feel stupid now than later.

A worked example

I ran the recipe on a real-shaped plan: "Launch a paid tier of my newsletter ($8/mo) in 3 weeks. I have 4,000 free subscribers and a ~40% open rate. Plan: announce next week, open paid signups in week 2, deliver the first paid-only issue in week 3. Goal: 200 paid subs in month one."

The ugly first attempt — when I gave a model just "what could go wrong with this plan?" with none of the constraints, here's the kind of thing it returned:

  • Low conversion: Your free subscribers may not convert to paid. Make sure the value proposition is clear.
  • Pricing: $8/mo may be too high or too low. Consider testing different price points.
  • Content quality: Paid subscribers expect more. Ensure the paid content justifies the cost.
  • Timeline: Three weeks is ambitious. Build in buffer time.
  • Churn: Some subscribers may cancel. Focus on retention.

Every line is true. Not one of them tells me what to do on Tuesday. "Make sure the value prop is clear" is not an action, it's a fortune cookie. This is the output that makes people think pre-mortems are a waste of time.

With the full recipe, the same model produced this (lightly trimmed):

Assume it's mid-July and you got 70 paid subs, not 200. Here's how:

# Failure mode Likely Impact Seen
1 The 40% open rate is inflated by Apple Mail auto-opens, so your real reachable audience is ~1,200, not 4,000 — and 200 paid was never on the table M H Late
2 You announce, get a spike of 50 signups in 48h, read it as "on track for 200," then signups flatline — because the people who'd pay paid immediately and there's no long tail H H Late
3 Week-3 paid issue ships late or thin because you're writing it the same week you're handling billing support and refund requests for the first time H M Early
4 Stripe/tax/payment setup eats 3 days you budgeted as "an afternoon," compressing content time M M Early
5 Free subscribers feel bait-and-switched when good content moves behind the paywall; unsubscribes spike and word-of-mouth (your only growth channel) dries up M H Late

Cheapest action this week for the top 3:

  1. (Inflated reach) Before you build anything, send one normal free issue with a single tracked link to a throwaway "what would you want in a paid tier?" page. Count unique clickers. That number — not 4,000 — is your real top of funnel. One issue, one afternoon, and it reprices your whole goal.
  2. (Flatline) Don't announce to everyone at once. DM or email your 30 most-engaged readers this week and ask them to pre-commit at $8. If fewer than ~10 say yes warmly, 200 is fantasy and you find out for free now.
  3. (Late/thin issue) Write the entire first paid issue THIS WEEK, before any of the billing/launch chaos exists. If it's done before you announce, failure mode #3 cannot happen.

The load-bearing assumption: "Enough of my free readers value this content at $8/mo specifically — not 'I like the free version.'" Go answer it: do the 30-reader pre-commit in action #2 before you write a line of launch copy. If that fails, nothing else matters.

That's the difference. The second version reprices the goal (200 was probably never real), and the #1 action is a half-day test that could save three weeks. Note it didn't just rank risks — it caught that the headline number was the actual problem.

(This is the output I got running the recipe, shown so you can judge the shape of it. It's not a captured or independently verified run — treat it as an illustration of what good looks like, not a benchmark.)

Where it breaks

It invents specifics that sound authoritative but are guesses. "Apple Mail inflates your open rate" is a real phenomenon, but the model doesn't know yours is inflated — it's a plausible story. Patch: treat every specific as a hypothesis to check, not a finding. The recipe's step-4 ("cheapest action to surface it") is your defense: each top risk comes with a way to verify it cheaply. Do the checks; don't take the stories as facts.

It defaults to the most common failure modes for the category, missing what's weird about your situation. It'll nail generic-startup risks and miss that your co-founder is leaving in week 2, because you didn't tell it. Patch: put the weird, specific, slightly-embarrassing constraints in the plan you paste. The deadline pressure, the skill you don't have, the stakeholder who hates this. The pre-mortem is only as good as what you confess into it.

It goes soft if your plan reads like a brag. Phrase the plan as "my brilliant launch strategy" and even a harsh prompt gets gentler. Patch: paste the plan as flat, neutral facts — numbers, dates, resources — with no adjectives. Let the model supply the skepticism; don't pre-load it with confidence.

Ten failure modes is paralysing if you treat them all as equal. Patch: that's exactly why the recipe ranks by Late + High-impact and then asks for only three actions. Ignore the bottom of the table. The ranking is the deliverable; the long list is just raw material.

It won't tell you the plan is fine. Ask for failure modes and you get failure modes — even for a genuinely solid plan. Patch: calibrate by reading the actions, not the risks. If the cheapest fixes are all easy and obvious, your plan is probably healthy. If the #1 fix is "actually, go verify the core assumption first," listen.

Cost & time

About 10 minutes: two to paste and run, eight to actually read it and do the one cheap check it surfaces. The check is the point — a pre-mortem you don't act on is just anxiety with a table.

Cost is a rounding error. One good model call with this prompt and a paragraph of plan runs roughly $0.02–0.04 on a current frontier model; on a free tier or a chat subscription you already pay for, it's effectively $0. The expensive version of this is the one where you skip it and find failure mode #1 in week three.

Run it on anything you can't easily take back. The first time it reprices a goal you were about to chase for a month, it pays for every lazy horoscope it ever gave you.

CAPTURED  A real captured run of this recipe — see the actual model output
captured 2026-06-17 · claude (local CLI) · committed to the public record
**Pre-mortem — Plan: "Convert my 4,000 free newsletter subscribers into 100 paying subscribers ($8/mo) in 8 weeks, solo, $0 ad spend, by adding a paid tier and emailing the existing list."**

It's August 12. The plan failed. Not in flames — you have 31 paying subscribers, growth has flatlined, and you're quietly telling yourself you'll "relaunch in the fall." Here's how you got here.

**Failure stories**

1. You announced the paid tier in week 1, got 22 sign-ups from your most loyal readers in 48 hours, and read that as traction. It wasn't traction — it was your entire pool of pre-sold fans converting at once. Weeks 2–8 added 9 people total because everyone left was a tougher sell, and you'd already used your one big announcement.
2. The free newsletter that built the list was opinion/commentary. The paid tier you pitched was "deeper analysis + a database." Those are different products. Free subscribers liked you for the thing you're now putting *behind* a paywall — so the paid pitch read as "pay to keep getting what was free," and reply-complaints made you soften the pitch.
3. Stripe + the paywall integration ate week 1 and half of week 2. By the time checkout worked, your launch email went out late and rushed, the urgency was gone, and you'd burned the freshest part of your motivation on plumbing.
4. You set $8/mo with no annual option. People who'd happily pay $80/year up front churned monthly after the first issue they skimmed. Three months of $8 < one $80 charge, and the monthly framing invited monthly reconsideration.
5. Open rates on the free list were 38%, which you assumed meant 4,000 engaged readers. Actually ~1,500 ever open, and the paid pitch only reached people who open — so your real addressable pool was 1,500, and 100/1,500 was a 6.7% conversion ask, not 2.5%.
6. You didn't email a paid-pitch more than twice because a few unsubscribes spooked you. The people who buy on email #4 never saw email #4.
7. The "database" you promised as the paid hook didn't exist yet at launch. You promised to build it live. By week 5 it was half-built, subscribers who paid for it asked where it was, and you spent weeks 5–7 building instead of selling.
8. Your free issues kept going out on schedule and stayed great, so there was no felt difference between free and paid. Nobody upgrades when the free thing already satisfies them.
9. You measured "100 subscribers" but never set a weekly checkpoint, so you didn't notice the week-2 flatline until week 5, when there was no runway left to change approach.

**Generics killed and replaced**
- ~~"Poor marketing"~~ → #1: front-loaded conversion masking a dead pipeline.
- ~~"Product-market fit issues"~~ → #2: paid product is a *different* product than the free one earned trust for.
- ~~"Technical delays"~~ → #3: payment plumbing on the critical path of week 1.
- ~~"Pricing wrong"~~ → #4: monthly-only pricing invites monthly churn.

**Ranked — late + high-impact at top**

| # | Failure | Likelihood | Impact | Visibility |
|---|---------|-----------|--------|------------|
| 1 | Loyalists convert early, pipeline is actually dead (#1) | H | H | **Late** |
| 5 | Real addressable pool is ~1,500, not 4,000 (#5) | H | H | **Late** |
| 2 | Paid tier is a different product than free earned (#2) | M | H | **Late** |
| 7 | Promised hook (database) doesn't exist at launch (#7) | M | H | Early |
| 4 | Monthly-only pricing → monthly churn (#4) | H | M | Late |
| 8 | Free product too good to upgrade from (#8) | M | M | Late |
| 6 | Stopped pitching after 2 emails (#6) | M | M | Early |
| 3 | Payment plumbing eats week 1 (#3) | M | M | Early |

**Top 3 — cheapest action this week**

1. **Dead pipeline (#1):** Before announcing anything, email 15 *non-loyalist* subscribers (people who open sometimes, never reply) one question: "Would you pay $8/mo for X? Why / why not?" If fewer than 3 say yes, your 100 number is fantasy and you redesign now, not in week 5. Cost: one email, one afternoon.
2. **Real pool is 1,500 (#5):** Pull your last-90-day unique-opener count today. Divide 100 by it. If the implied conversion rate is above ~5%, the target is wrong, not the execution — fix the target before you build anything. Cost: one analytics query.
3. **Wrong product (#2):** Take your 3 best-performing free issues and your planned paid pitch, and write one sentence: "Free is for ___, paid is for ___." If you can't make those two different without making the free version worse, you don't have a paid product — you have a paywall. Cost: 20 minutes and honesty.

**The one assumption everything rests on:**

> Do at least ~75 of your subscribers want a *different, deeper* thing than what you already give them free — enough to pay for it — as opposed to merely liking the free thing you already give them?

If the answer is no, pricing, plumbing, pacing, and emails are all irrelevant. Go answer it this week with action #1, before you build anything.