The line between agent tasks that work and ones that flop

A year in, the agent tasks that pay off and the ones that flop split on a single question: did I write the context down, or is it still in my head?

Written down → it works:

1. Point it at a folder of call transcripts; it writes the FAQ from the repeats.
2. A research skill that runs before every intro call.
3. Turn a 40-minute session into a one-page brief I can forward.
4. rg the workspace for "where did we decide X" instead of me digging for it.
5. Draft the changelog from the week's commits and sessions.

Still in my head → it flops:

• Anything needing taste I never bothered to articulate.
• "Simple" tasks it gets confidently wrong — AI is spiky, it'll nail a brutal synthesis and then misread a calendar screenshot.

The leverage comes from writing the context and the taste down so the agent can actually reach them. Prompt-crafting barely matters next to that.