The anatomy of a good prompt

A model can only work with what the prompt gives it. Ask it to summarise this incident and nothing else, and the recorded reply is a polite request for the incident: there is nothing to summarise. That is the problem every prompt solves, and the six plates in the film are the six kinds of information a request can carry: the task, who it is for, the material to work from, the limits, an example of a good answer, and the exact shape the answer must take.

The film builds one real request, a summary of a checkout outage for the on-call channel, one plate at a time, and every card shows what the model actually replied. Nothing on the cards was written by hand: each version of the prompt was sent once to `google/gemini-3.8-flash` at temperature 0 on 11 September 2026, and the replies are shown verbatim, including their Markdown.

With the context plate the model writes 96 words of accurate bullet points, a report for nobody in particular. Naming the audience makes it longer, 101 words, and adds a recommendation about concurrent index creation that the incident log never mentions: told who is reading, it tried to be helpful beyond the evidence. The constraints plate, at most 60 words and only what the log shows, cuts the reply to 53 words and the guess goes away.

The surprise is the example. A single sample summary turned the reply from prose into JSON on its own, a bigger change than any other plate, because an example shows the model the pattern rather than describing it. But the reply arrived wrapped in a Markdown code fence and without a follow_up key, so a bot that runs JSON.parse on it fails. Only when the format plate names all six keys does the reply parse first time.

That is the failure mode worth remembering: a reply that looks right to a person and breaks the program that reads it. The explorer holds the removal experiments too. Without the example plate the reply still parsed, but the follow_up field held a fact from the log rather than an action. Without the format plate it parsed but had five keys. Without the constraints plate the speculative advice came back.

The honest caveat is that this is one request, one model and one recording per prompt, so the size of each effect is specific to this case: another model, another task or another run at a higher temperature will move differently. The shape of the lesson travels well, though. When a program consumes the reply, the example and the format are not optional, and a plate like the audience, which mattered a lot for a person reading, barely changed a reply whose shape was already pinned down.