A TASK_SPEC you can check
1000 split three ways doesn't come out even, and if nobody wrote down a rule for the remainder, the agent picks one itself. Using a restaurant bill, we build a TASK_SPEC together with Claude Code: where the task ends, where the extra cent goes, and which numbers prove the result right.
Running a large task: context and checkable stages
A locale catalog with a list, search and a filter is easy to ask for in one sentence, and then it takes a while to find what broke. We split it into stages you can open and check one by one, and save the state in PROGRESS.md so you can pick up where you left off after /clear.
Four types of tasks for an agent
The banner button leads to /old/ instead of /campaign/, and the quickest fix is exactly the wrong one: the address lives in the data. We go through four types of tasks - a new feature, a bug fix, a refactoring and research - each with its own first step and its own check.
How to choose a Claude Code permission mode
Asking "don't change anything" is still just asking: whether edits are allowed is up to the permission mode, and only the indicator tells you which one is on. We lay out the Claude Code modes from plan to bypassPermissions and see when it's easier to leave auto for a stricter rung.
GO TO FULL VERSION