How to Use Claude Subagents for Real Work Tasks (Step-by-Step)
Claude Subagents are one of the most powerful features in Claude Pro, and also one of the most skipped. Most users either don't know the feature exists or assume it's for developers. It's not. Once you understand the basic pattern, you can use Subagents for real work tasks in under ten minutes.
This guide walks through exactly how Subagents work, when to use them, and a concrete example you can try today.
What Claude Subagents Actually Are
A Subagent is Claude working through a multi-step task where it breaks the work into smaller pieces and handles them in sequence. Instead of giving Claude one big prompt and hoping it does everything at once, you're giving it a task structure, and Claude figures out the steps.
Think of the difference between asking an assistant "write me a market analysis" versus saying "here's a research task: first gather the key data points, then organize them by theme, then write a summary for an executive audience." The second approach produces better output because the task has structure.
Subagents formalize that structure. Claude runs through phases, passing the output of one phase as input to the next.
When Subagents Are Worth Using
Not every task needs a Subagent. If you're writing a quick email or asking a one-shot question, basic prompting is faster.
Subagents make sense when:
- The task has a clear sequence of steps that depend on each other
- Each step would benefit from Claude focusing on it separately rather than doing everything at once
- The final output needs to synthesize multiple types of work (research, then analysis, then writing)
- You've tried a single large prompt and the output feels thin or disorganized
Common use cases include content research and drafting, competitive analysis, data summarization across multiple inputs, and generating structured reports from raw notes.
Step-by-Step: Using Claude Subagents Without Code
You don't need the API for this. You can run Subagent-style workflows directly in the Claude interface by structuring your prompts correctly.
Step 1: Define the Final Output First
Before writing a single prompt, know what you want at the end. A vague goal produces a vague result. Be specific.
Bad: "Help me analyze this market." Good: "I want a 400-word executive summary identifying the three biggest competitive threats in [market], based on the information I'll provide."
Write that final output definition down. You'll use it at the end.
Step 2: Break the Task Into 3 to 5 Phases
Most tasks have a natural structure. A content research workflow might look like:
- Extract the key facts and claims from the source material
- Identify gaps or questions the source doesn't answer
- Organize the findings into themes
- Write the final summary
Each phase should have one clear job. If you can't describe what a phase does in one sentence, split it further.
Step 3: Write a Prompt for Phase 1
Start with a prompt that handles only Phase 1. Give Claude the context it needs, the source material or the task setup, and ask only for the Phase 1 output.
Example for a competitive analysis:
"I'm going to give you a set of notes about three competitors. Your job right now is only to extract the key claims each competitor makes about their product. Don't analyze or compare yet. Just list the claims by company.
Here are the notes: [paste your notes]"
Run this. Read what Claude returns. If it's good, move to Phase 2.
Step 4: Pass the Output Into Phase 2
In the same conversation, give Claude the Phase 2 instruction. Reference what it just produced.
"Good. Now using those claims, identify the three areas where the competitors overlap and the two areas where they're making completely different bets. Still no conclusions yet, just the pattern."
This is the key mechanic: each phase builds on the last. Claude holds the context from the previous turn, so you don't need to re-paste everything.
Step 5: Continue Through Each Phase
Work through each phase you defined in Step 2. Keep each prompt focused on that phase's job only. Resist the urge to ask Claude to do two phases at once. The separation is what makes the output good.
If a phase returns something unexpected or weak, you can ask Claude to redo it with a correction before moving forward. This is much easier than trying to fix a combined output after the fact.
Step 6: Run the Final Phase
Once you have the structured output from all intermediate phases, give Claude the final instruction, using your output definition from Step 1.
"Now using everything we've built, write the 400-word executive summary I described at the start. The audience is a VP who will read this in two minutes and needs to understand the three biggest threats."
The final output will be significantly better than if you'd asked for it in a single prompt at the beginning. Claude had time to build the structure before writing the narrative.
A Template You Can Copy
Here's a reusable Subagent template for any research-to-report workflow:
Phase 1 prompt:
"I'm working on [task]. I'll give you the raw material now. Your only job in this step is to extract [specific type of information]. Don't draw conclusions yet.
[paste your material]"
Phase 2 prompt:
"Using what you just extracted, identify [patterns, themes, gaps, or comparisons]. Still no final conclusions."
Phase 3 prompt (optional):
"Now organize those [patterns/themes] into a structure that will work for [your final output format]."
Final phase prompt:
"Now write the final [report/summary/analysis] using everything we've built. Here's the spec: [your output definition from Step 1]."
How to Save and Reuse This Workflow
Once you've run a Subagent workflow that works, save it. The prompt structure from each phase is reusable on the next similar task.
ClaudeSkill has a Personal Prompt Vault where you can save your own workflows and organized prompts with tags and notes. Free users get up to 10 vault entries, which is enough to store the workflows you'll actually use regularly. If you find yourself building more, the Pro plan at $12/month removes that limit.
ClaudeSkill also has structured recipes for Subagents and other Claude features, organized by role. If you're a marketer or analyst, there are recipes built specifically for your work tasks rather than generic examples. The Foundation tier is free and worth trying before you build everything from scratch.
You can explore the full recipe library at claudeskill.vercel.app.
Common Mistakes to Avoid
Combining phases. The most common mistake is asking Claude to do two phases at once because it feels faster. It isn't. The output suffers.
Skipping the output definition. If you don't know what you want at the end, Claude doesn't know either. Spend 60 seconds writing it down before you start.
Not reading phase outputs before continuing. Each phase is a checkpoint. If the Phase 2 output is wrong, fixing it before Phase 3 is much easier than untangling it at the end.
Starting a new conversation between phases. Keep everything in one conversation. Claude needs the context from earlier turns to make the later phases work.
Frequently Asked Questions
Do I need to pay for Claude Pro to use Subagents? Subagent-style prompting works in any Claude plan. The technique is just a way of structuring your conversation. You don't need a special feature toggle.
How is this different from just writing a long prompt? A long prompt asks Claude to do everything at once. A Subagent workflow asks Claude to do one thing at a time, in sequence. The outputs are usually more structured and accurate because Claude can focus on each step fully.
Can I use Subagents for writing tasks, not just research? Yes. A common writing workflow: Phase 1 gathers or organizes the raw ideas, Phase 2 creates an outline, Phase 3 drafts section by section, Phase 4 edits for tone and clarity. Each phase produces a cleaner result than a single "write this article" prompt.
How long does a Subagent workflow take? A three-phase workflow for a standard task usually takes 10 to 20 minutes total, including your reading and correction time between phases. That's still faster than doing the same work manually.
Where can I find pre-built Subagent recipes for my role? ClaudeSkill has role-specific Subagent recipes for marketers, analysts, writers, ops managers, and developers. The Foundation recipes are free. Intermediate and Advanced Subagent workflows are on the Pro plan at $12/month.