Search across all documentation pages
Ten practical habits for delegating tasks and writing descriptions Claude can act on, grouped into three areas: judging what to delegate, writing the description itself, and closing the loop once you see the result.
Because they happen in sequence during real use: a delegation judgment call comes first, a description follows once that call is made, and closing the loop happens after Claude responds; grouping them this way mirrors how the habits actually get used.
Stating the actual goal, not just the task, tends to have the single biggest effect on output quality, since it is the piece most often left implicit in a bare request.
Treat it as a signal about the delegation decision itself, not only the prompt, and reconsider whether the task actually belongs in a human-led or partially delegated category.
Only when the request actually mixes a judgment call with drafting or research; a request that is purely one or the other does not need to be split.
Because the audience determines what needs explaining and what can be assumed as already known; without it stated, Claude has to guess, and that guess is often wrong for a specific reader.
A vague constraint like "keep it short" leaves the actual limit to guesswork; a concrete constraint like "under 150 words" gives Claude something specific to check the output against.
An example communicates format, structure, and voice more precisely and faster than a paragraph of description trying to capture the same thing in words.
Because Claude's output is typically well-written and confident in tone regardless of whether the underlying content is fully correct, which makes an off-target or wrong answer easy to miss without deliberately reading for it.
It means identifying which specific piece, context, a constraint, or the stated goal, was actually missing or unclear, and fixing that piece directly instead of guessing broadly at a full rewrite.
Yes; a task type like drafting a public statement is not fixed as safe or risky, its stakes shift with context such as timing and audience, so the delegation call should be made per instance, not by category alone.
No, this list is a quick-reference summary; the full Delegation and Description pages explain the reasoning and edge cases behind each habit in more depth.
Stack versions: Written against the Claude model lineup current as of ~June 2026 - Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5 (the default), and Claude Haiku 4.5. Model names, pricing, and product features move quickly - verify current specifics at platform.claude.com/docs before relying on them.
Reviewed by Chris St. John·Last updated Jul 16, 2026