Interactive example · from Working with Coding Agents
The third time
Same as last week: normalize the new samples to -14 LUFS, rename them to the naming scheme, move them into the library folders.
That's the third time we've run this exact sequence by hand: loudness scan, rename, file. I can just do it again, or this is the signal to build something.
Pick the agent's next move, then see where it leads:
Week five: half my new samples are way quieter than the rest. What happened?
In that run I passed -24 instead of -14 to the normalizer, a typo in a hand-typed command, and nothing checked it. Each manual repetition re-rolls the dice on every flag; this time one came up wrong, silently.
Manual repetition has variance, and variance compounds silently. "Faster to just do it" is true exactly once. By the third repeat, the two minutes are an illusion priced in future mistakes.
Written and run on the real batch: 14 samples normalized, renamed, filed. Output attached. The target level is a flag, so the day you want -16 for a client, nothing gets rewritten. Next week this is one command, for you or for me.
And when I add the spectral-tagging step later?
Separate small tool, piped after this one. They compose through files: that's the point of keeping each one small.
The repetition became a tool: small, policy-free (the threshold stays with the caller), proven on real input the same session. Every future run costs one command and zero variance.
Three weeks later: how do I make it just normalize one folder? And why does the watcher keep re-processing old files?
The one-folder case needs profile overrides in the config, and the watcher bug is an interaction between two flags... the menu system is now the thing we maintain instead of the thing that works.
The app in disguise: interacting options, a config file, state. A maintenance surface nobody asked for, wrapped around what wanted to be three small composable tools.