Kieran Hunt

Simon Willison wrote this interesting piece about using an llm in a script’s shebang. Here’s an expansion of what he wrote but for OpenCode, written to a .sh file and given execuable permissions:

#!/usr/bin/env -S opencode run --dangerously-skip-permissions "follow the attached prompt" -f
Generate a limerick about a pelican riding a bicycle.

Running ./pelican.sh:

> build · global.anthropic.claude-opus-4-7

A pelican perched on a bike,
Pedaled off on a lengthy-beaked hike.
With wings tucked in tight,
He was quite the sight,
Rolling through town—what a strike!

It still supports all of the parameters that OpenCode does:

./pelican.sh --model amazon-bedrock/global.anthropic.claude-sonnet-4-5-20250929-v1:0

And iterating:

./pelican.sh --continue "now make more metal 🤘"

A pelican shredding on chrome,
Tore through the streets like a dome.
With talons on steel,
He thrashed every wheel,
Headbanging his way back to home!

It’s not example 1 to 1 with Simon’s implementation. llm -f inlines file content as a prompt fragment; OpenCode’s -f attaches a file. Functionally equivalent here, but opencode requires a non-empty positional [message..], hence the "follow the attached prompt" directive. For Simon’s tool-call examples, OpenCode’s tools are always available — you just need --dangerously-skip-permissions (or -i to allow approvals).

It’s a pretty useful way of persisting a pre-canned prompt. Maybe as part of a cron job or in a agent skill’s scripts directory.