Letting an assistant run the channel
What changes when the thing deciding what to post is a model with tools, and what has to stay fixed so that it is safe.
An assistant with access to your workspace can do the whole loop: read what happened, decide it is worth a post, write the headline, choose the channel and the language, and schedule it. This is not a demo any more. It is a Tuesday.
What the assistant should decide
Everything that is judgement over language:
- Is this event worth a post at all, or is it noise.
- Which of three angles is the interesting one.
- How to say it in nine words.
- Which channels care, and in which languages.
- Whether it goes out now or at six.
What the assistant should never decide
Everything that is the brand:
- The type, the colours, the logo, the spacing, the safe area.
- Whether a headline that does not fit gets published anyway.
- What a chart’s axis starts at.
- Whether the legal line at the bottom is optional.
The distinction matters more than it sounds. A system where a model can adjust the layout to make its copy fit is a system that will, eventually and with the best intentions, publish something off-brand at three in the morning. A system where the composition is fixed and the copy is validated against it will simply refuse, and say which field was too long.
The boring safety property
A refusal is a better outcome than a clipped headline, and it is only available if the constraint lives in the template rather than in a prompt. That is why fields are validated before a render is ever started, and why the error names the field.
An assistant that can be told no is an assistant you can leave running.
