Before hope edits the bar
Startup Kill Criteria — When to Stop Before You Sink More Build Time
You already know how the week should end if you are honest. The hard part is writing that ending down before the compliments, the sunk hours, and the “one more sprint” talk. Kill criteria are that page: a tripwire you can execute without renegotiating with hope.
Last updated: September 28, 2026
Direct answer
What are startup kill criteria?
A kill criterion is a Stop contract you write before the next week of work: an observable signal, a threshold you would accept as decisive, a calendar window, and a named action — Kill, Pivot, or Continue. You evaluate later evidence against that page. You do not invent a friendlier bar after people were polite. Gary Klein’s project pre-mortem (Harvard Business Review, September 2007) asks a team to treat the plan as already failed and write why. This page is the next sheet: tripwires attached to those failure modes so Stop is executable when the evidence arrives. Scores do not kill ideas. A failed critical assumption, measured against criteria you wrote while you were still honest, does.
Key takeaways
What to remember
- Write the tripwire before recruitment. A bar invented on Sunday is a recap, not a criterion.
- Four fields, nothing vaguer: signal + threshold + window + action. “Give it more time” is not a field.
- A score is a flashlight. It can point at a weak dimension. It cannot fire a kill switch.
- Kill and Pivot are different actions. Write both in advance so mixed evidence cannot be stretched into “keep building the same thing.”
- After a criterion fires, the cheapest next move is a field test on the rewritten claim — not a new factory.
Why it matters
Hope is a terrible editor
I have watched founders treat “we need more time” as if time were evidence. Time is what you spend while the bar quietly moves. CB Insights’ 2026 analysis of 431 VC-backed companies that shut down since 2023 is useful here as a warning about symptoms versus roots — not as a personal forecast. “Ran out of capital” leads their list at about 70 percent, and they are explicit that this is often the final cause of death. The more telling patterns, among the 385 companies with identifiable reasons, include poor product-market fit (about 43 percent), bad timing (about 29 percent), and unsustainable unit economics (about 19 percent). Many shutdowns cited more than one reason, so those shares can add to more than 100 percent. Vague “90 percent of startups fail” slogans help even less. U.S. Bureau of Labor Statistics data on private-sector establishments — not a VC-backed census — shows a steeper early drop, then a slower decline: a TED note dated January 12, 2024 reports that 34.7 percent of establishments born in March 2013 were still operating in March 2023, about one-third at ten years. The same TED table shows survival falling 20.4 percentage points in the first year (to 79.6 percent) and 50.6 percent still operating at five years (March 2018). The useful lesson is not a slogan. It is that a lot of dying happens early, and “keep going because I already started” is how a founder becomes a later exit without ever writing Stop.
Vocabulary
Four terms this page uses on purpose
These match how the rest of Yibud talks. If a sentence here uses one of these words, it means this.
Kill criterion (Stop contract)
A pre-written tripwire: signal + threshold + window + action. You write it before the evidence window opens. You honor it when the window closes. A risk list without a dated action is not a criterion.
Signal
An observable behavior you could photograph or paste into a log: a payment, a second booked session, a sent file, a parent email that actually went out, or the absence of a weekly incident. “Interest,” “good conversations,” and “they liked the idea” are not signals.
Threshold
The count or yes/no you would accept as decisive if a stranger brought you the log. “Fewer than 3 of 8 qualified people did the action” is a threshold. “If it doesn’t feel promising” is a mood.
Window and action
The window is calendar dates, not “a quarter” or “until we know.” The action is what you will do when the window closes: Kill the current one-liner, Pivot to a rewritten person or job, or Continue into the next test. If the action is blank, hope will fill it in.
Not the same page
Kill criteria vs pre-mortem vs assumption vs the week-after decision
Four Yibud pages sit next to each other. Mixing them up wastes a week — or three months of code.
Pre-mortem — invent the failure history
A pre-mortem treats the one-liner as already dead and ranks the ways it died. That hour produces failure modes. It does not yet say what evidence, by which date, means you will stop. Kill criteria are the tripwires you attach after the ranking.
Startup idea pre-mortem (30–45 minutes) →Critical assumption — kill one claim
That page names the single claim that would make the rest of the plan irrelevant if false, then runs a cheap week against Continue / Refine / Stop lines. This page is the Stop contract those lines are written against: what evidence means Stop versus Pivot, dated before anyone talks to you.
Critical assumption first — name, falsify, decide →Validation decision — after a week of evidence
That page is the call once interviews, a smoke test, or a payment log already exist. This page is written before that emotional investment, so the later call is executable instead of renegotiated.
When to continue, pivot, or kill after evidence →A mid-band overall score is a flashlight, not a kill switch. It ranks relative risk in the inputs you typed. It does not decide Stop. What a validation score actually means →
The sitting
How to write kill criteria in one sitting
Do this after a pre-mortem ranking, or after a report names a critical assumption — not after you have already spent the week explaining the vision. Set a timer. Leave the build closed. If you have a co-founder, each of you writes a draft alone, then you keep the stricter bar.
- 1
Name the lethal assumption, not the product
One sentence: the named person will [observable action] because [mechanism]. If you only have a one-liner, run the pre-mortem hour first and take Rank 1. If you already have a Yibud report, start from the named critical assumption — as a sentence to test, not as a verdict.
- 2
Write the signal you could log without a story
Pick one observable: a payment, a second session, a sent file, a message they actually used, or “no weekly incident in the last Tuesday they described.” If the only thing you can write is “they were interested,” you do not have a signal yet. Tighten the person or the action until a no is possible.
- 3
Write the threshold a stranger would accept
A number or a clean yes/no. “3 of 8 qualified people do the action” or “zero paid second cycles.” Do not write a threshold you would never honor (“if nobody on earth wants this”). Do not write a threshold that requires a factory (“if we cannot hit 10,000 users”).
- 4
Date the window on a calendar
Start date and end date. Fourteen days from first outreach is a window. “Give it the summer” is not. If life closes the window early, you reschedule the dates in writing — you do not silently keep building and call it the same test.
- 5
Write Kill, Pivot, and Continue as actions
Kill: what you will stop building, and what notes you will keep. Pivot: the one rewrite allowed — person or job, not a feature. Continue: the next cheapest test, not a backlog. If Pivot and Continue read the same, you have not written Pivot.
- 6
Date the page and put it where you cannot lose it
Same sitting. Sign it. If you later want a kinder bar, that is a new page with a reason — not an edit in the margin after the first compliment. Then run the cheapest field test the assumption needs. Steve Blank’s Customer Development is the posture after the contract exists: leave the building. It is not a reason to skip the contract.
What you leave with
The Stop Contract (copy this)
If the sitting produced a mood, you did not finish. The page should have one filled contract and three action lines. Fill the brackets. Leave nothing as “users” or “soon.”
The tripwire
Assumption: [named person] will [action] because [mechanism]. Signal: [observable]. Threshold: [N of M / yes-no]. Window: [start date]–[end date].
If any bracket still says “people,” “traction,” or “later,” the contract is not written. Mechanism examples: no weekly pain, will not pay the planned price, cannot be reached in the named channel, already has a good-enough workaround.
The three actions
Kill if [threshold missed]: stop [named build], keep [notes], do not open [the tempting factory]. Pivot if [mixed pattern]: rewrite [person or job] the same day, write a new contract. Continue if [threshold met]: next test is [one cheap field action].
Pivot is a rewrite of the sentence, not a new feature on the same sentence. Continue is one next test, not permission to “just start.”
After the tripwire fires
If Kill or Pivot fires, the cheapest next test is [conversation / one-page / manual job] on the rewritten claim — not a new repo.
Blank’s Customer Development belongs here: leave the building on the new sentence. A criterion that fires and then becomes a six-month rebuild is a criterion you did not honor.
The fork
Kill, Pivot, or Continue — written before the evidence
The contract does not replace the field week. It decides what the week is allowed to mean. The critical-assumption page is how you run the week. The validation-decision page is how you read a pile of evidence you already have. This page is why Sunday cannot invent a kinder story.
Kill
The threshold was missed on the lethal assumption. Stop the current one-liner. Write what you will not build. Keep the notes. A different sentence gets its own contract. Do not “kill” by quietly renaming the same product.
Pivot
The evidence is mixed in a way you named in advance — they want the job, not the price; the person is real, the channel is not. Rewrite the person or the job the same day. Write a new Stop contract. Adding a feature is not a pivot.
Continue
The threshold was met on a signal a stranger could log. The next move is the next cheapest test, not a feature backlog. Continue is not “we are right.” It is “this sentence survived this window.”
Worked example
One contract you can copy (illustrative)
The week below is illustrative — names, prices, and counts are invented so you can see the setup. It is not a conversion benchmark and not a Yibud report. Copy the method, not the numbers.
The sentence, written before outreach
“A $39/month inbox that drafts Sunday-night parent emails for elementary teachers who already write them by hand.” The tempting next step is a template library. The lethal assumption is: a named teacher who wrote parent emails last Sunday will send a draft we wrote, or pay $39 for two weeks of manual drafts, because the Sunday-night pile is already a weekly incident.
The tripwire
Signal: a qualified teacher sends a real parent email they actually used, or pays $39 for two weeks of drafts done by hand. Threshold: fewer than 3 of 8 qualified teachers do either. Window: 14 calendar days from the first outreach message. Qualified means: they taught this term and wrote at least one parent email last Sunday. Compliments in the thread do not count.
The three actions, written the same sitting
Kill if fewer than 3 send or pay: stop the inbox product; keep the Sunday-night notes; do not open the template library “so the next eight have something to click.” Pivot if they describe the Sunday pile and want help, but refuse email drafts — rewrite toward a 20-minute Sunday checklist or a live call, then write a new contract. Continue if 3 or more send or pay: the next test is a second two-week cycle at the same price, still by hand.
What must not happen in the window
Do not widen “elementary teachers” to “anyone who emails parents.” Do not count “I would use that” as a send. Do not extend day 14 because the eighth person has not replied — silence is a no on reach, and reach may be the lethal mode. If you cannot find eight qualified teachers in fourteen days, that can be the result: rewrite the contract around reach, or Kill.
Illustrative close: write the contract before the first Slack DM. If day 14 arrives and two paid, three complimented, and three never replied, that is Kill or a Pivot you already named — not a reason to start the template library “because some people were warm.”
Where Yibud fits
Use the free validator as a flashlight, then write the contract yourself
Yibud is a free, no-signup startup idea validator. Scores come from a deterministic rule engine, not from a language model guessing success. Optional AI text, when it is used, only polishes prose. It does not invent the number. A Startup MRI report can name a weak dimension and a critical assumption. That is a starting flashlight. This page is the Stop contract you write against that sentence. If you already have a report, start the contract with the named assumption — not with the overall score. Then run the week yourself. The method stands alone if you never open the analyzer.
Analyze my idea →Common mistakes
What usually wastes the contract
Writing the bar after the compliments
Sunday’s “we got good signal” is not a criterion. If Kill / Pivot / Continue were blank on Monday, you have a recap. Hope is an excellent editor of empty pages.
Using the score as a kill switch
A low overall does not kill the idea. A high overall does not save it. The score ranks relative risk in the inputs you typed. A failed critical assumption against a dated threshold is what stops the build.
A threshold you would never honor
“If nobody in the world wants this” is not a tripwire. Neither is “if we cannot become a unicorn.” Write a number you would actually stop for, on a person you can actually find this window.
A window with no calendar
“Give it a quarter” and “until we know” are how the bar moves. Dates. If life interrupts, reschedule in writing. Do not silently keep building and call it the same test.
A signal that cannot produce a no
Interest, warmth, and “good conversations” cannot fail. Ask about last Sunday, ask for the file, ask for the $39, or watch whether the weekly incident exists. Politeness cannot trip a wire you refused to string.
Sources
Where these ideas come from
- CB Insights, “The top 9 reasons startups fail” (report analyzing 431 VC-backed shutdowns since 2023) — Used as a warning about symptoms versus roots, not as a personal success forecast. CB Insights reports “ran out of capital” at about 70 percent and calls it often the final cause of death. Among the 385 companies with identifiable reasons, more telling patterns include poor product-market fit (about 43 percent), bad timing (about 29 percent), and unsustainable unit economics (about 19 percent). Many companies cited more than one reason. This page does not invent other percentages.
- U.S. Bureau of Labor Statistics, The Economics Daily, “34.7 percent of business establishments born in 2013 were still operating in 2023” (January 12, 2024) — Used to push back on vague “90 percent of startups fail” slogans. The TED note reports that 34.7 percent of U.S. private-sector establishments born in March 2013 were still operating in March 2023 — about one-third at ten years. The accompanying table shows the steepest drop in year one (20.4 percentage points, to 79.6 percent) and 50.6 percent still operating at five years (March 2018). These are establishments, not a VC-backed startup census.
- U.S. Bureau of Labor Statistics, Business Employment Dynamics — Establishment Age and Survival Data — The program behind the TED note. Survival declines steeply early, then more slowly. Cited as the primary government series, not as a rule that any one idea should be killed on a schedule.
- Gary Klein, “Performing a Project Premortem” (Harvard Business Review, September 2007) — The business write-up of treating a plan as already failed, then generating reasons. Kill criteria attach dated tripwires to those failure modes. Klein does not claim the exercise predicts startup outcomes.
- Steve Blank, Customer Development — Used only after a criterion fires: the cheapest next test is a field hypothesis you leave the building to check. Blank is not the source of the Stop contract, and this page does not treat Customer Development as a sample-size rule.
In one paragraph
Summary you can quote
Startup kill criteria are a Stop contract written before the next week of hope: an observable signal, a threshold you would accept, a calendar window, and a named Kill / Pivot / Continue action. A pre-mortem invents the failure history; a critical-assumption week tests one claim; the validation-decision page reads evidence you already have. This page makes Stop executable so you do not renegotiate the bar after compliments. Scores do not kill ideas. Failed critical assumptions against pre-written criteria do. Vague “90 percent fail” slogans are a poor substitute for a dated threshold — U.S. establishment survival data show a steep early drop, then a slower decline, and they are not a personal forecast. A Yibud score can point at a weak dimension. The contract still has to be written by the founder.
FAQ
Questions founders actually ask
How is a kill criterion different from a pre-mortem?
A pre-mortem invents a failure history and ranks modes. A kill criterion is the tripwire you evaluate later: signal, threshold, window, action. Do the hour first if you only have a one-liner. Write the contract next so Stop is not a mood.
Can a Yibud score kill the idea?
No. The score is a structured read of the inputs you typed. It does not interview anyone or collect a payment. A failed critical assumption against a bar you wrote before the window is what stops the build. Skip the contract only if you already have behavioral evidence against the same sentence.
What if I miss the window because life happened?
Reschedule the dates in writing and keep the same threshold. Do not silently keep building and call it the same test. If you cannot reach the named person inside a window you can actually work, reach may be the lethal assumption — rewrite the contract around that, or Kill.
How do I write Kill and Pivot in advance without boxing myself in?
Pivot is one rewrite you name now — person or job — for a mixed pattern you can already imagine. Kill is what you do if the lethal signal does not appear. If every mixed outcome still says “keep the same product,” you have not written Pivot. You have written Continue twice.
Do I need an account, and where does Yibud fit?
You do not need an account. The method on this page stands alone. Run the free analyzer if you want a flashlight on weak dimensions before you write the contract. The engine is deterministic. Optional language-model text only polishes prose. Bring the named assumption back here and fill the brackets.
Write the Stop contract before the next sprint
Yibud scores weak dimensions with a deterministic engine. You write signal, threshold, window, and action while you are still honest, then decide Kill / Pivot / Continue against that page — not against Sunday’s story.
Generate a Yibud report →