Cited from real sources 6 min read Written by Arun Agrahri Updated October 2026

A decision framework by Annie Duke

Annie Duke's Kill Criteria for Founders Deciding to Quit

Kill criteria are the conditions you agree to in advance for stopping or pivoting. Annie Duke wrote Thinking in Bets and Quit. She found that a premortem on its own changes almost no one's plan. Its real job is to surface the signals you would see if this failed. Attach an action to each signal while you can still think straight about it.

Best for: Founders with too many projects who need to agree in advance when to stop or pivot one
Price: Free to read. To apply it to your own situation, Gavel has a free plan; Plus is $19/mo.

The framework

The premortem is not the deliverable

Most teams treat a premortem as a planning exercise. Imagine the project failed. List the reasons. Feel sober. Then go build the thing you were always going to build. Duke's observation is that this is what happens: run a premortem and people do not change their plan much. The insight evaporates, because nothing forces it to survive contact with a decision.

Her reframe is to treat it as a signal generator. The output you want is not a revised plan, it is a list of things you would observe on the way down. Each becomes a kill criterion: a measurable condition meaning stop or pivot, written while you have no stake in the answer.

It has to be structural, not intentional. Once you launch, the biases pointing at continue outrun your resolve. Sunk cost is the loudest. Duke's inversion: waste is forward-looking. You cannot get back what you already spent; the real waste is the next six months poured into something you know is dead.

things go on way too long even like they're over budget and you've blown the timeline and when you finally shut it down you realize you should have done it a lot earlier

Duke on how slow teams are to quit after launch Watch at 66:21

So the framework is not think harder about risk. It is convert the risk you already imagined into named signals, and pre-commit to what you will do when each one appears.

How to apply it

How do you set kill criteria that actually fire?

Seven moves, drawn from how Duke runs the exercise and from the failure modes she names.

1

Run the premortem as a signal hunt, not a plan review.

Prompt the team with a dead future: it is six months from now and this failed. Ask what early signals they saw, not what they would change.
2

Have people write independently before anyone talks.

Duke uses nominal group technique for this. Opinions gathered in the room produce conformity, and the quietest useful signal never gets said.
3

Convert each signal into a measurable condition.

Not 'if traction is weak' but a metric and a date. Vague criteria are the most common reason the exercise does nothing.
4

Attach an action to every criterion.

A threshold with no committed response is a discussion. Decide now whether hitting it means kill, pivot, or a specific test that resolves it.
5

Instrument the leading indicators.

If nobody can see the signal in real time, it will not fire. Wire criteria to something you already watch, or the trigger arrives as hindsight.
6

Hold quit reviews on a fixed cadence, separate from ops.

Inside a status meeting, optimism bias dominates and every miss gets explained away. The review needs its own slot and its own question.
7

Honor the pre-commitment when it triggers.

The predictable failure is renegotiating at the moment the criteria fire. Duke's warning: quitting on time will almost always feel like quitting too early.

A worked example

Duke describes running this with a sales team. Every rep got the same prompt: you won a lead through an RFP. You worked it six months. The deal is now dead. Looking back, what were the early signals? The list was long. Three became kill criteria, each with an action attached.

Someone wrote the RFP with a competitor in mind. Action: ask them whether they are working with that competitor and how far along it is. Kill or proceed on the answer.
The customer will not take a demo and only wants to talk about price. Action: kill it. The team decided this pattern means the customer is checking a box so they can beat someone else down on price.
No decision maker in the room after the first few meetings. Action: offer executive alignment at the next meeting. If the answer is no, kill it.

Note what this does for the rep. Walking away invites the question of why you stopped pursuing the deal. A pre-agreed criterion turns that from a judgment call into a policy the team already signed off on.

Boundary conditions

When it works, when it fails

Works best when

The commitment is long enough that sunk cost will accumulate before you get proof either way
You can name leading indicators, not only the outcome you are waiting on
You set them before launch, before anyone's identity attaches to the answer
Someone other than the project owner is in the room when the team reviews the criteria

Fails when

The criteria are too vague to be actionable, which is the most common version of the mistake
Quit reviews happen inside the standard operating meeting, where optimism bias wins
The team renegotiates the criteria at the exact moment they are about to trigger
No one instruments anything, so the signal only becomes visible in the postmortem

Knowing about a bias does not protect you from it. Duke treats sunk cost like a visual illusion: you can explain it. You can demonstrate it. You can understand it. You can still fall for it. Only a constraint built before the temptation enters the room changes what you do.

instead of just kind of hoping that when I see those signals I actually act rationally which is a hope that it will not come true

Duke on why intention is not a plan Watch at 69:22

Where operators disagree

Duke locates the problem in the individual. Resulting, motivated reasoning and hindsight bias drive decision errors. She fixes decisions by fixing personal process. The systems thinkers disagree about where the leverage sits. Donella Meadows and Peter Senge argue structure produces behavior. Feedback loops, incentives and delays matter too. Blaming individual judgment misses the cause. Cassie Kozyrkov sits closer to Duke: decision hygiene, set criteria before you see the data.

Both levels operate at once. Kill criteria are a personal-process intervention. They work best when the structure backs them. That means a review cadence on the calendar, and incentives that do not punish the rep who walked away from a dead deal.

Kill criteria pair with the frameworks that generate the signal. Sean Ellis's 40 percent product-market fit test is a ready-made threshold you can commit to before you run it. Garry Tan's launch-the-jankiest-version rule is the other half: ship early enough that the signal arrives while quitting is still cheap.

The sources

Where Duke discusses this

65:59

Timestamp

Lenny's Podcast · This will make you a better decision maker

Why a premortem exists: setting up kill criteria

Duke reframes the premortem's real purpose as generating kill criteria, the signals that mean it is time to pivot or stop.

67:57

Timestamp

Lenny's Podcast · This will make you a better decision maker

The sales-team example, three criteria with actions

An RFP written with a competitor in mind, a customer who only discusses price, no decision maker in the room, each wired to an action.

32:00

Timestamp

Lenny's Podcast · This will make you a better decision maker

The Illusion of Long Feedback Loops

Challenging the notion of long feedback loops, Duke argues that decision-makers can actively shorten them by identifying and tracking predictive signals.

40:00

Timestamp

Lenny's Podcast · This will make you a better decision maker

The Difficulty of Quitting and Learning from Failure

Duke discusses the psychological biases that make quitting difficult, such as sunk cost fallacy and identity protection.

Useful? Pass it to a founder six months into something they already know is dead.

Want the full playbook?

Get 457 strategic thinking frameworks.

89 frameworks 125 rules 235 heuristics & principles 11 operators

From Annie Duke, Donella Meadows, Ray Dalio, and 8 more. Drop one .md into Claude, Cursor, or ChatGPT. Your AI cites practitioners, not guesses.

See the pack

Instant .md download · One-time purchase · No subscription

Related frameworks

Shipping

Weekly Review

David Allen

Once a week, get every list and project current so your head can let go. An hour or two buys the rest of the week.

Shipping

Habit Stacking

James Clear

Pin a new habit to one you already do every day: after I make my coffee, I will meditate for 60 seconds.

Shipping

Time Blocking

Cal Newport

Run the workday from a schedule, not a to-do list. Give every minute a job in advance, redraw the plan when it breaks, and stop at the shutdown.

Shipping

Levels of AI Autonomy

Amjad Masad

Grade a coding tool by how long it runs before it needs you. Level one is a language server, level five is a thousand agents.

The premortem is not the deliverable

How do you set kill criteria that actually fire?

Run the premortem as a signal hunt, not a plan review.

Have people write independently before anyone talks.

Convert each signal into a measurable condition.

Attach an action to every criterion.

Instrument the leading indicators.

Hold quit reviews on a fixed cadence, separate from ops.

Honor the pre-commitment when it triggers.

When it works, when it fails

Works best when

Fails when

Where Duke discusses this

Get 457 strategic thinking frameworks.

Know when the next expert lands.