sharing session[ murphy's law for ai ]

anything that can go wrong will go wrong.ai just gets there faster.

murphy's law as a mindset for discovery, mvps and prototyping, now that making is cheap.

krishna

[ disclaimer ] i practice this. i have not mastered it. this is me sharing the parts i want to be more consistent at.

origin[ music · meaning · work ]
hukummurphy
hukum murphykafin sulthan · songwriter & performer

i came for the music.
i stayed for the mindset.

01 / a quieter mindmusic, when i needed it.
02 / a familiar sound, made hereto my ears: french pop meets shibuya-kei. anak negeri.
03 / a question i could bring to workwhat could go wrong, and how do we prepare?
from listening to product discoveryat ipc · many perspectives. one question still open.
(00)the moment[ murphy's law for ai ]

making got cheap. being right did not.

effort to make a draft,a mock, a prototype cost when a wrong thing reaches customers the optimism gap before ai now

illustrative shape, not measured data.

  • a PRD, a mock, a working prototype: each one a prompt away.
  • so everyone feels fast, and everyone feels right.
  • that is exactly when murphy shows up.
(01)the symptom[ murphy's law for ai ]

ai sounds just as sure when it is wrong.

prd: one-tap reorderai draft
  1. goal: members reorder a past purchase in one tap.

  2. impact: lifts repeat purchase by 23%.

    ✕ invented number. no source, no baseline.

  3. flow: member taps reorder, the order is placed.

    ✕ happy path only. out of stock? price changed? address deleted?

  4. integration: real-time sync with the loyalty engine.

    ✕ which engine? nobody checked it exists.

  5. edge cases: all handled.

    ✕ none are listed.

  6. status: ready for development.

    ✕ no non-goals. no owner. no undo.

sample draft written for this slide. every line reads fine. five of six would cost us later.

the framework[ murphy's law for ai ]

four habits, borrowed from things that break.

a pessimist's toolkit for an optimistic moment.

01murphy's law
expect failure

expect the wrong way.


in ai work

count the steps. each one can fail.

02reka peluang
edge cases

map the ugly paths first.


in ai work

ask for the failure list before the feature list.

03buffers
margin

leave slack where the bad week lands.


in ai work

a second reader, capped review, time to check.

04recovery
resilience

make the failure survivable.


in ai work

know your undo before you ship.

(02)the law[ murphy's law for ai ]

ai adds steps for free.
murphy charges for every one.

  • each step 99% right
  • each step 95% right
  • each step 90% right
chance at least one is wrong steps or claims in the work 40% a 10-step agent chain 64% a 20-claim spec
(03)the origin[ murphy's law for ai ]

murphy's law was a design brief, not a mood.

1949, a rocket-sled test for the us air force. a set of sensors turned out to be installed the wrong way round. the takeaway, as the story is usually told: if it can be done wrong, someone will do it wrong.

the engineering answer was not 'be more careful'. it was: make the wrong way impossible.

✕ ✓

wrong way: blocked

right way: fits

sim card: the cut corner fits one way only.

no wrong side

usb-c: designed so there is no wrong way in.

only fits its hole

poka-yoke: toyota's word for mistake-proofing.

in ai work

  1. 01constrain before you generate.
  2. 02give it a template, a schema, the real system names and the non-goals.
  3. 03a form, not a blank page.
(04)reka peluang[ murphy's law for ai ]

imagine it already failed. then list why.

1935: a new bomber prototype crashed on its evaluation flight. a control lock had been left on. the fix was not a better pilot. it was a checklist.

today's version is the pre-mortem: before launch, the team assumes it already failed and writes down why.

pre-mortem: gary klein

pre-mortemreka peluang
one-tap reorder shipped. it failed. why? ✕item out of stock ✕price changed since last order ✕saved address deleted ✕voucher already used ✕double tap, two orders ✕payment method expired

in ai work

  1. 01ask the ai for the failure list before the feature list.
  2. 02then turn every branch into a test case or a non-goal.
(05)buffers[ murphy's law for ai ]

slack looks like waste until the bad week.

demand spike + late truck.the buffer takes the hit, not the customer. stock on hand weeks reorder point safety stock

retail keeps safety stock because demand and delivery both wobble. on a good week it just sits there. on a bad week it is the business.

in ai work

  1. 01a second model attacks the draft cold.
  2. 02review capped at two rounds, then a human decides.
  3. 03time booked for checking, not only for making.
(06)recovery[ murphy's law for ai ]

design the fall, not just the climb.

01 / past case

✕rope cut safety catch holds ✓
1854 / the rope breaks. the catch holds.otis made the failure survivable.

02 / building with ai

independent review · model b

buildermodel a skepticchallenge assumptions? verifiertest the result humanrelease checkpoint reviserollback

a different reviewer brings another perspective. it still needs checks.

multimodal: images · audio · videocoding: implement · test

the mindset[ murphy's law for ai ]

optimistic about the goal.
pessimistic about every step.

before anything ai made reaches a customer:

  • ✓count the steps. each one is a chance to be wrong.murphy's law
  • ✓write the non-goals and ugly paths before the prompt.reka peluang
  • ✓never let the author grade its own work.buffers
  • ✓know your undo before you ship.recovery

making is cheap now. checking is the job.

disclosure[ murphy's law for ai ]

this deck was made 100% by ai.

human direction. ai production. human judgment.

my workflow roles krishnadirection naquuuubotcoordinate naquuuu-curatortaste librarianresearch scribestory builderbuild skepticchallenge verifiercheck humandecision
v1 · too many examplesv2 · too much showing offv3 · shaped by feedback

by slide 6's math, something in here is probably wrong.

ai draft · check anything important before using it.

human loop[ murphy's law for ai ]

now the human loop. your turn.

krishna ai you v4
  1. 01what is wrong in this deck?
  2. 02which failure have we not mapped in our own work?
  3. 03where would a buffer have saved us?

[ q&a ]what you find goes into v4.

← notes
[ catatan pembicara ]