A launch asset is ready only when every product claim in it can survive a direct question: “What evidence lets us say that?” If I cannot trace a sentence to observed product behaviour, the sentence is generated copy, however polished the video looks.
In 1999, NASA’s Mars Climate Orbiter approached Mars after a journey of hundreds of millions of kilometres. The spacecraft disappeared because one engineering team had produced thruster data in pound-force seconds while another system expected newton seconds. NASA’s Mars Climate Orbiter Mishap Investigation Board documented the mismatch after the loss.
The numbers passed through the system. They looked usable. Their meaning had not survived the handoff.
I thought about that failure while watching a launch video I had considered ready.
The sentence that changed the review
The edit had reached the comfortable stage. The timing worked. The product appeared clean on screen. The narration carried the viewer from the problem to the result without slowing down to expose every technical decision underneath it.
Then one sentence landed with complete confidence.
I knew what the product did. I had seen the feature run. Still, I could not explain exactly what evidence supported the full claim as spoken.
That distinction mattered. A working feature can support several accurate statements and still fail to support the strongest available version. The model may generate an answer, but does it do so consistently across realistic inputs? The automation may complete a task, but what happens when a required field is missing? The demonstration may succeed, but did we choose the example because it represented normal use or because it avoided the difficult case?
The video gave no sign of those unresolved questions. Good production had compressed them into certainty.
I stopped reviewing the edit as a viewer and began reviewing it as the person who would have to defend each sentence after launch.
A generated claim can sound observed
AI makes this problem easier to create.
A founder can move from a rough script to a convincing voiceover, product sequence and finished edit in one afternoon. That speed is useful. It also removes several moments where doubt used to surface naturally.
When a team wrote every line, recorded several takes and rebuilt awkward scenes, weak claims had more chances to become uncomfortable. Generation reduces that friction. The first plausible sentence can travel from prompt to script to narration before anyone identifies whether it describes tested behaviour, intended behaviour or hoped-for behaviour.
Those categories sound similar during production. They become very different when a customer buys because of the sentence.
The review therefore had to move below wording. I needed to find the source of the claim.
Had I watched the product produce this result?
Had someone else tested it under conditions a customer would recognise?
Was the sentence describing the current build, an internal prototype or the roadmap?
Could I show the evidence without explaining away an important caveat?
These were product questions disguised as copy edits. That is common in early-stage teams. Marketing reveals ambiguity that a demo can hide because a public claim forces the team to choose one meaning.
The same issue appears when an engineering assumption crosses a boundary without its context. I wrote about that failure pattern in The Missing Assumption in Friday’s CAD Pull Request, and What Monday Inherits. A launch script is another handoff. Product knowledge enters one side; a buyer’s expectation leaves the other.
The defence test
I now use a simple review condition: someone must own the defence of every consequential claim.
Ownership means more than approving the wording. The owner should be able to name the product state, test, customer observation or operating constraint behind it. If the evidence supports a narrower sentence, the sentence gets narrower. If the claim depends on a condition, the condition belongs in the copy or the scene. If nobody can locate the evidence, the line leaves the video.
This can feel conservative when runway is short and competitors speak with fewer caveats. Yet an unsupported sentence borrows conversion from the future. The first customer who tests its edge case will collect the debt through support, rework or distrust.
There is also a difference between uncertainty and weakness. “The system drafts a response for review” may sell less excitement than “The system handles your customer replies.” It may sell the product that actually exists.
That accuracy helps the right buyer make the right decision. It also gives the team a cleaner product target. If we want the stronger sentence later, we can name what must become true before we use it.
What I would ship
I would keep the video only after building a small claim ledger beside the script. Each important sentence would have an evidence source, an owner and any condition that changes its meaning. No elaborate approval process. A plain document is enough.
Then I would ask one person who did not write the script to challenge the most confident line. Their task would be specific: identify what a reasonable buyer could believe after hearing it, then compare that belief with the product’s current behaviour.
NASA’s investigation into the Mars Climate Orbiter found more than a unit mismatch. It found that the processes intended to detect the problem had failed to do so. That is the bridge to a founder reviewing an AI-generated launch asset: the dangerous claim may travel cleanly through every production step because each step assumes somebody else verified its meaning.
The final review should break that chain.
Pause on the sentence that sounds easiest to repeat. Open the product. Reproduce the result. Write down the conditions. Then decide whether you are shipping a claim you can defend or a sentence the tools made easy to believe.
Comments
No comments yet.