Split the job: 5 prompts that use subagents

Split the job: 5 Claude Code prompts that send work to subagents

A subagent is a helper Claude sends off to do one part of a job in its own separate workspace, then report back. These are 5 prompts for splitting real work across subagents, how to watch them run, and when splitting costs more than it saves.

What you need

• Claude Code, on a Pro, Max, Team or Enterprise plan or a Console account.

• A job with 3 or more parts that don't depend on each other.

• For web research: in Manual mode, Claude Code asks before it searches the web. A subagent's request shows up in your main session with the subagent's name on it.

How it works

Anthropic's subagents page: "Each subagent runs in its own context window with a custom system prompt, specific tool access, and independent permissions." The reading and searching stays in the subagent's window, and only its summary comes back to your conversation.

They run in the background. From the Claude Code changelog, v2.1.198: "Subagents now run in the background by default, so Claude keeps working while they run and is notified when they finish." You keep talking to Claude while they work.

You ask in plain words. The docs' own example:

Research the authentication, database, and API modules in parallel using separate subagents

Say "in parallel" and "one subagent each", and say what each one should hand back. The docs add that this "works best when the research paths don't depend on each other."

While they run:

• Running subagents show in a panel below the prompt.

• When one finishes, its row leaves the panel and the footer shows /tasks to see subagents for 30 seconds. Run /tasks during those 30 seconds and press Enter on it to read its transcript.

• Press Ctrl+B to send a running task to the background.

• If a subagent needs your OK, the prompt names it. Approve, or press Esc to deny that one action without stopping the subagent.

Forks: a helper that already knows the conversation

"A fork is a subagent that inherits the entire conversation so far instead of starting fresh." Use one when a fresh subagent would need too much background to be useful. Start it yourself with /subtask and the task:

/subtask draft a reply to the supplier email we discussed, using the points we agreed above

It runs in the background while you keep working, and its result arrives as a message in your conversation. Because it shares the main session's cache, the docs say a fork "costs less than a fresh subagent for tasks that need the same context."

The 5 prompts

Run each one in Claude Code. Change the words in [brackets]. Each prompt keeps the reports short, because every report lands back in your conversation.

1. Research 3 vendors

Research these 3 vendors in parallel, one subagent each: [vendor A], [vendor B], [vendor C].

Each subagent checks the vendor's own website for: what they offer for [my need], what it costs if they publish it, the limits that matter for [my team size], and how to reach sales. Each one returns at most 150 words plus the URLs it used. If a fact isn't on a page it can cite, it writes "not found".

When all 3 are back, put them in one table, one row per vendor, and tell me which one to call first and why.

2. Review 3 documents

Review these 3 documents in parallel, one subagent each: [file 1], [file 2], [file 3].

Each subagent reads its document in full and returns: a 3-line summary, every date, deadline and amount with the page or section it's on, and anything that conflicts with [my standard terms / the brief I've described]. Quote the exact wording for anything it flags. Keep each report under 200 words.

Then compare the 3 reports and list every place the documents disagree with each other.

3. Check a folder of files

Check every file in the folder [folder name]. Split the files into 3 groups and give each group to its own subagent.

Each subagent opens every file in its group and checks: [what to check, for example: has a date, has a signature line, the total matches the line items]. Don't change any file. Each one returns only the files that fail a check, with the file name and the reason.

Then merge the 3 lists into one, sorted by file name, and tell me how many files were checked in total.

4. Compare options

I need to choose between [option A], [option B] and [option C] for [the goal].

Send one subagent to make the strongest case for each option. Each one returns: the case in 5 bullet points, the biggest risk, and what would have to be true for it to be the right choice. Each one works only from [the files I've named / public sources it can cite], and marks anything it infers as an inference.

When all 3 are back, put them side by side and tell me which one you'd pick, with your confidence: high, moderate or low.

5. Draft and critique

Write 3 drafts of [the email / the proposal section / the post] in parallel, one subagent each. Draft 1 takes [angle 1], draft 2 takes [angle 2], draft 3 takes [angle 3]. Each one uses only the facts in [file or notes].

When the 3 drafts are back, start one more subagent that isn't a fork, and give it only the 3 drafts and this test: [what the reader must understand and do after reading]. It scores each draft against the test and names the weakest sentence in each.

Then show me the best draft with the critic's fixes applied, and the critic's notes underneath.

The critic in prompt 5 starts fresh, so it reads the drafts the way a new reader would. That's the point of giving it only the drafts.

The mistake to avoid

Splitting small jobs. A subagent "sends its own requests, which count toward the same usage limits as your main conversation," and "each subagent spends tokens of its own while it runs." A subagent that isn't a fork also starts fresh and may need time to gather context. For a quick, targeted change, or anything that needs back-and-forth with you, the docs say to stay in the main conversation. Split when the parts are independent and each one means a lot of reading you don't need to see.