Daily ICM Tip: If a stage needs a paragraph to explain what it does, it's actually two stages.

The trap everyone falls into when they're new to this is building a folder like '02_analyze_and_draft',

and then wondering why the output is mushy. The agent read the research, formed an opinion, and committed to prose all in one pass, so you can't see where the thinking went sideways. There's no edit surface between "what it concluded" and "how it wrote it up." When the draft is wrong, you're re-running the whole thing instead of fixing the one bad handoff.

"One stage, one job" isn't a tidiness rule, it's a debugging rule. The folder boundary is the review gate. Cramming two jobs into one stage doesn't save you a step.

look at any CONTEXT.md and read its job out loud. If you naturally say "and" like, "it summarizes the sources and picks the angle" that "and" is a folder boundary you haven't cut yet.

Every 'and' in a stage's job description is a review gate you're throwing away.

1 comment