dax sends phone screenshots straight into an opencode session
The screenshot is the prompt, and Peter Steinberger says the pattern is already old news.
dax says he has started taking a screenshot on his phone and sending it straight into an opencode session. The example he gave was a screenshot of DMs on X about scheduling a call, forwarded to the agent with the instruction "schedule this". He called the flow amazing.
take a screenshot on your phone and send straight to an opencode session
Peter Steinberger replied "welcome to 2025", which is the whole counterpoint in three words. Nobody is claiming this is a new capability. What dax is reporting is that it has become his default way to hand an agent a task from a phone.
welcome to 2025
Why the screenshot is the interesting part
The usual phone workflow is transcription. You read the thing on screen, you summarise it into a prompt, you type the summary, and the agent acts on your compression of the original rather than the original. A screenshot skips that step entirely. The image carries the names, the times, the back and forth about the call, and the agent pulls what it needs out of the picture. The typed part collapses down to two words of intent.
That also changes what the instruction has to do. "Schedule this" is not a specification. It only works because the screenshot is the specification, and the agent is expected to read the context rather than be told it. On a phone, where typing is the bottleneck, that is the difference between queueing a task and not bothering.
This is one engineer describing one session, not a measured result. dax did not say what happened next, whether the call got scheduled correctly, what tooling opencode used to do it, or how often the pattern works. The sources do not say. Treat it as a workflow report, not evidence that screenshot prompting is reliable.
It fits a pattern dax has been building
It is also consistent with how he has been working. He has described running agent sessions that he checks in on rather than babysits, including one pinned session per project that farms work out to subagents. A phone screenshot is the smallest possible input to that kind of setup. You are not sitting down to work, you are dropping something into a session that is already running.
For working engineers, the copyable idea is narrow and cheap to test. Next time you would have typed a paragraph summarising something already on your screen, send the screen instead and write the verb. If your agent harness accepts images, there is nothing to install. If it does not, that is the thing worth fixing before any of the rest of this applies.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.