Skip to content
Leonidov.DEV

Premortem Skill: Assume the Launch Failed First

Business
Premortem Skill: Assume the Launch Failed First

A release with a green checklist can still be rolled back three weeks later, because nobody asked what happens if the client’s team never opens the settings screen. Every item was verified and every test passed, and neither fact covered that question.

Code: Pre-Mortem-Skill, GPL-3.0 licensed.

A risk review usually produces four items, three of which are already on the plan, plus a note that says “monitor after launch”. The failure that lands is rarely on the list. That gap is the reason the premortem runs before anything that costs real money, and the reason it runs from a file: a process that depends on my mood does not survive a deadline.

The sentence that does the work

“It failed. Why?”

That is the whole trick. The project is already dead, the room is not defending it, and the question becomes an explanation instead of a prediction. People who would never say “I think this will fail” will happily explain how it did.

Ask a model to critique a launch plan and the default answer is a polite summary of strengths, then a short list of risks phrased so gently that nothing needs changing. Asking “this plan failed, explain how” produces a different document: named dependencies, a support queue with no owner, a migration that assumed an empty database. The question, not the model, does the work.

What the premortem skill does differently

The skill is a Markdown file with a seven-step protocol, and the constraints matter more than the steps.

Boring failures are mandatory. At least two of the generated causes have to be mundane: timeline slippage, motivation decay, a budget moved to another project. Without that rule a model reaches for dramatic causes, and dramatic causes are rarely what kills a small project.

The horizon is explicit. Three months for a launch, twelve for a hire. An open horizon makes every cause plausible and no cause specific.

Likelihood and impact get scores from one to five, and the scores are visible in the room. That sounds bureaucratic until the first disagreement: a failure rated two comes back as five from the person who knows the vendor contract.

Every failure gets an early warning sign and a mitigation that fits in the current week. A mitigation that needs a quarter of planning is not a mitigation, it is a second project.

The report ends with a gap check that names what the analysis probably missed. A premortem with a stated blind spot is worth more than one that claims to be complete.

The step everyone skips

The protocol asks for one commitment: a single action with a specific day. It is the first thing to disappear when the report is long and the fixes look obvious.

Then the obvious fix happens in the following month, late. Writing “rotate the staging certificate before the 14th” into the transcript takes ten seconds and saves a Saturday.

Where it does not help

A premortem on a vague idea produces vague failure. If the plan cannot be stated with a date and a success condition, the exercise has nothing to bite.

It is also useless on a decision already made. Explaining how a signed contract failed is called regret, and it does not change the contract.

And it will not save a project that fails because of something nobody in the room controls. It will, however, surface how much the plan depends on that something, which is often the useful part.

Why a file instead of a habit

A habit needs a memory and a free hour. A file needs neither, and it keeps the same seven steps even when the team is tired and wants to move fast.

That is the pattern I keep coming back to: if a process matters, write it where the agent can read it and let the checklist argue with me. The same reasoning sits behind how a missing spec burns a project and behind why a business without goals does not grow, which is the same problem at a larger scale.

Background: Gary Klein, Performing a Project Premortem. The skill is on GitHub under GPL: Pre-Mortem-Skill.