@Claude TaskThe craft at the centre of your mission: a delegation prompt that decomposes cleanly.
ClaudeTag "break[s] its task down into stages and then work[s] through them in turn."[ref §2] That decomposition is only as good as the task you hand it. This is the lesson where "learn ClaudeTag" and "write better agent prompts" become the same skill.
Because delegation is async and visible, you can't clarify mid-thought the way you would in a back-and-forth. The task has to stand on its own. Four parts make it decomposable:
| Part | Question it answers |
|---|---|
| Objective | What outcome, in one sentence? |
| Success criteria | How will we both know it's done and correct? |
| Constraints | What must it not do / what bounds cost & risk? |
| Output shape | What form should the result take? |
Same intent, two very different results:
@Claude how's the account doing?Vague objective, no success criteria, no output shape. Claude guesses; the thread fills with clarifying questions; you're back to synchronous chat.
@Claude Objective: give me a 1-line health read on the live FTMO
account for the last 24h.
Success: covers current DD vs the soft limit, open positions,
and whether any EA went silent.
Constraints: read-only tools only; don't fire any control; if data
is stale >90s say so instead of reporting it.
Output: one Slack message, prefixed with a UTC timestamp.Now it decomposes cleanly: pull DD → check positions → check EA heartbeat → format one line. No round-trips.
@Claude task is just your global rules, scoped to one job.
Decomposition has limits. A task spanning ten loosely-related goals produces a muddled plan; one atomic instruction wastes the async model. Aim for one objective that naturally splits into a handful of stages — the sweet spot ClaudeTag is built for.
Say the objective aloud. If it needs the word "and" more than once to join unrelated goals, split it into separate tasks. If it has no internal stages at all ("what's the current DD?"), it's a quick query, not a delegation — just ask. The delegation sweet spot is a single outcome with 2–6 implied steps.
Which element most reliably stops the clarifying-question ping-pong?
Success criteria tell Claude when to stop and what "correct" means, removing the main reason it pauses to ask. Greetings and spend caps don't reduce ambiguity.
"@Claude fix trading and also redo the finance report" should be:
Two unrelated goals (different domains, likely different channels/identities) muddle the plan. Split them — and they probably belong to different Claude identities anyway (Lesson 4).
You have a repeatable four-part template — objective, success, constraints, output — for any delegation, and a size heuristic. Paste it above any @Claude task, or any subagent prompt, and the decomposition takes care of itself.